Dileep George: Brain-Inspired AI
AI 与机器学习生物与进化技术与编程音乐与艺术心理与人性
📋 章节目录
暂无章节信息
🔑 关键词
brainmodeldonableinferenceneuralconnectionsnetworklearningmodelslanguagetryingcorticalinterestingvisualcorrecttextcortexdoinghuman
💬 精彩语录
"very top and then zoom in. Okay. So, one important thing, constraint that went into the model is that"
非常顶部，然后放大。好的。因此，模型中的一件重要的约束是
— Dileep George （38:58.880）
"like the human brain, and I wouldn't say capture is a solved problem. We have cracked the fundamental"
就像人脑一样，我不会说捕获是一个已解决的问题。我们已经破解了根本
— Dileep George （53:27.040）
"have hypotheses about how those lateral connections are supposedly contributing to visual processing."
对这些横向连接如何促进视觉处理有假设。
— Dileep George （19:16.160）
"A. And locally, just when I look at just that patch of the image, it looks like an A. But when I look"
A. 在本地，当我只看那块图像时，它看起来像 A。但是当我看时
— Dileep George （1:00:03.200）
"mentioned, bringing the ideas together in a unique way? Is there something there? Is there some value"
提到过，以独特的方式将这些想法结合在一起？那里有什么东西吗？有没有一些价值
— Dileep George （1:12:36.080）
🎙️ 完整对话（1217 条）
Lex Fridman (00:00.000)
The following is a conversation with Dilip George, a researcher at the intersection of
以下是与交叉学科研究员 Dilip George 的对话
Lex Fridman (00:05.360)
Neuroscience and Artificial Intelligence, cofounder of Vicarious with Scott Phoenix,
神经科学和人工智能，Vicarious 与 Scott Phoenix 的联合创始人，
Lex Fridman (00:10.880)
and formerly cofounder of Numenta with Jeff Hawkins, who's been on this podcast, and
以及 Numenta 的前联合创始人，与杰夫·霍金斯 (Jeff Hawkins) 共同创立了这个播客，并且
Lex Fridman (00:16.800)
Donna Dubinsky. From his early work on hierarchical temporal memory to recursive cortical networks
唐娜·杜宾斯基。从他早期对分层时间记忆的研究到递归皮层网络
Lex Fridman (00:23.520)
to today, Dilip's always sought to engineer intelligence that is closely inspired by the
直到今天，迪利普一直致力于设计智能，其灵感来源于
Dileep George (00:29.600)
human brain. As a side note, I think we understand very little about the fundamental principles
人类的大脑。顺便说一句，我认为我们对基本原理了解甚少
Dileep George (00:35.760)
underlying the function of the human brain, but the little we do know gives hints that may be
人类大脑功能的基础，但我们所知甚少，暗示可能
Dileep George (00:41.600)
more useful for engineering intelligence than any idea in mathematics, computer science, physics,
对于工程智能来说，比数学、计算机科学、物理学中的任何想法都更有用，
Lex Fridman (00:46.960)
and scientific fields outside of biology. And so the brain is a kind of existence proof that says
以及生物学以外的科学领域。所以大脑是一种存在证明
Dileep George (00:53.120)
it's possible. Keep at it. I should also say that brain inspired AI is often overhyped and use this
这是可能的。坚持下去。我还应该说，受大脑启发的人工智能经常被夸大，并使用它
Dileep George (01:01.040)
fodder just as quantum computing for marketing speak, but I'm not afraid of exploring these
就像营销中的量子计算一样，但我并不害怕探索这些
Dileep George (01:08.000)
sometimes overhyped areas since where there's smoke, there's sometimes fire.
有时过度炒作的地区，因为有烟的地方有时会发生火灾。
Dileep George (01:13.680)
Quick summary of the ads. Three sponsors, Babbel, Raycon Earbuds, and Masterclass. Please consider
广告的快速摘要。三个赞助商：Babbel、Raycon Earbuds 和 Masterclass。请考虑
Dileep George (01:20.400)
supporting this podcast by clicking the special links in the description to get the discount.
通过单击描述中的特殊链接来支持此播客以获得折扣。
Dileep George (01:25.760)
It really is the best way to support this podcast. If you enjoy this thing, subscribe on YouTube,
这确实是支持这个播客的最佳方式。如果您喜欢这个东西，请在 YouTube 上订阅，
Dileep George (01:31.440)
review it with five stars on Apple Podcast, support on Patreon, or connect with me on Twitter
在 Apple Podcast 上以五颗星评价它，在 Patreon 上提供支持，或者在 Twitter 上与我联系
Dileep George (01:36.400)
at Lex Friedman. As usual, I'll do a few minutes of ads now and never any ads in the middle that
在莱克斯·弗里德曼。像往常一样，我现在会做几分钟的广告，中间不会有任何广告
Dileep George (01:42.400)
can break the flow of the conversation. This show is sponsored by Babbel, an app and website that
可能会打断谈话的流畅性。该节目由 Babbel 赞助，这是一个应用程序和网站
Dileep George (01:48.960)
gets you speaking in a new language within weeks. Go to babbel.com and use code LEX to get three
让你在几周内用一种新语言说话。访问 babbel.com 并使用代码 LEX 获取三个
Dileep George (01:54.480)
months free. They offer 14 languages, including Spanish, French, Italian, German, and yes, Russian.
免费几个月。他们提供 14 种语言，包括西班牙语、法语、意大利语、德语，当然还有俄语。
Dileep George (02:03.040)
Daily lessons are 10 to 15 minutes, super easy, effective, designed by over 100 language experts.
Dileep George (02:10.560)
Let me read a few lines from the Russian poem Noch ulytse fanar apteka by Alexander Bloc
Lex Fridman (02:18.160)
that you'll start to understand if you sign up to Babbel.
Dileep George (02:34.720)
Now I say that you'll only start to understand this poem because Russian starts with a language
Lex Fridman (02:41.440)
and ends with vodka. Now the latter part is definitely not endorsed or provided by Babbel
Lex Fridman (02:47.600)
and will probably lose me the sponsorship, but once you graduate from Babbel,
Dileep George (02:51.760)
you can enroll in my advanced course of late night Russian conversation over vodka.
Dileep George (02:56.320)
I have not yet developed an app for that. It's in progress. So get started by visiting babbel.com
Lex Fridman (03:02.800)
and use code LEX to get three months free. This show is sponsored by Raycon earbuds.
Dileep George (03:09.360)
Get them at buyraycon.com slash LEX. They become my main method of listening to podcasts,
Dileep George (03:14.960)
audiobooks, and music when I run, do pushups and pull ups, or just living life. In fact,
Dileep George (03:20.880)
I often listen to brown noise with them when I'm thinking deeply about something. It helps me focus.
Dileep George (03:26.880)
They're super comfortable, pair easily, great sound, great bass, six hours of playtime.
Dileep George (03:33.920)
I've been putting in a lot of miles to get ready for a potential ultra marathon
Lex Fridman (03:38.080)
and listening to audiobooks on World War II. The sound is rich and really comes in clear.
Lex Fridman (03:45.760)
So again, get them at buyraycon.com slash LEX. This show is sponsored by Masterclass.
Dileep George (03:52.640)
Sign up at masterclass.com slash LEX to get a discount and to support this podcast.
Dileep George (03:57.840)
When I first heard about Masterclass, I thought it was too good to be true. I still think it's
Dileep George (04:02.400)
too good to be true. For 180 bucks a year, you get an all access pass to watch courses from
Dileep George (04:08.160)
to list some of my favorites. Chris Hatfield on Space Exploration, Neil deGrasse Tyson on
Dileep George (04:13.360)
Scientific Thinking and Communication, Will Wright, creator of SimCity and Sims on Game Design.
Dileep George (04:19.280)
Every time I do this read, I really want to play a city builder game. Carlos Santana on guitar,
Dileep George (04:26.240)
Garak Kasparov on chess, Daniel Nagano on poker and many more. Chris Hatfield explaining how rockets
Dileep George (04:32.640)
work and the experience of being launched into space alone is worth the money. By the way,
Dileep George (04:38.160)
you can watch it on basically any device. Once again, sign up at masterclass.com to get a discount
Lex Fridman (04:43.600)
and to support this podcast. And now here's my conversation with Dileep George. Do you think
Dileep George (04:50.960)
we need to understand the brain in order to build it? Yes. If you want to build the brain, we
Dileep George (04:56.400)
definitely need to understand how it works. Blue Brain or Henry Markram's project is trying to
Dileep George (05:04.160)
build a brain without understanding it, just trying to put details of the brain from neuroscience
Dileep George (05:11.920)
experiments into a giant simulation by putting more and more neurons, more and more details.
Lex Fridman (05:18.160)
But that is not going to work because when it doesn't perform as what you expect it to do,
Dileep George (05:26.560)
then what do you do? You just keep adding more details. How do you debug it? So unless you
Dileep George (05:32.720)
understand, unless you have a theory about how the system is supposed to work, how the pieces are
Dileep George (05:37.360)
supposed to fit together, what they're going to contribute, you can't build it. At the functional
Dileep George (05:42.400)
level, understand. So can you actually linger on and describe the Blue Brain project? It's kind of
Dileep George (05:48.560)
a fascinating principle and idea to try to simulate the brain. We're talking about the human
Dileep George (05:56.080)
brain, right? Right. Human brains and rat brains or cat brains have lots in common that the cortex,
Dileep George (06:03.600)
the neocortex structure is very similar. So initially they were trying to just simulate
Dileep George (06:11.200)
a cat brain. To understand the nature of evil. To understand the nature of evil. Or as it happens
Dileep George (06:21.040)
in most of these simulations, you easily get one thing out, which is oscillations. If you simulate
Dileep George (06:29.120)
a large number of neurons, they oscillate and you can adjust the parameters and say that,
Dileep George (06:35.200)
oh, oscillations match the rhythm that we see in the brain, et cetera. I see. So the idea is,
Dileep George (06:43.280)
is the simulation at the level of individual neurons? Yeah. So the Blue Brain project,
Dileep George (06:49.040)
the original idea as proposed was you put very detailed biophysical neurons, biophysical models
Dileep George (06:59.200)
of neurons, and you interconnect them according to the statistics of connections that we have found
Dileep George (07:06.320)
from real neuroscience experiments, and then turn it on and see what happens. And these neural
Dileep George (07:14.240)
models are incredibly complicated in themselves, right? Because these neurons are modeled using
Dileep George (07:22.080)
this idea called Hodgkin Huxley models, which are about how signals propagate in a cable.
Lex Fridman (07:28.240)
And there are active dendrites, all those phenomena, which those phenomena themselves,
Dileep George (07:34.000)
we don't understand that well. And then we put in connectivity, which is part guesswork,
Dileep George (07:40.960)
part observed. And of course, if we do not have any theory about how it is supposed to work,
Dileep George (07:48.960)
we just have to take whatever comes out of it as, okay, this is something interesting.
Lex Fridman (07:54.800)
But in your sense, these models of the way signal travels along,
Dileep George (07:59.440)
like with the axons and all the basic models, they're too crude.
Dileep George (08:04.320)
Oh, well, actually, they are pretty detailed and pretty sophisticated. And they do replicate
Dileep George (08:12.960)
the neural dynamics. If you take a single neuron and you try to turn on the different channels,
Dileep George (08:20.800)
the calcium channels and the different receptors, and see what the effect of turning on or off those
Dileep George (08:28.400)
channels are in the neuron's spike output, people have built pretty sophisticated models of that.
Lex Fridman (08:35.360)
And they are, I would say, in the regime of correct.
Dileep George (08:41.120)
Well, see, the correctness, that's interesting, because you mentioned at several levels,
Lex Fridman (08:45.680)
the correctness is measured by looking at some kind of aggregate statistics.
Dileep George (08:49.440)
It would be more of the spiking dynamics of a signal neuron.
Lex Fridman (08:53.200)
Spiking dynamics of a signal neuron, okay.
Dileep George (08:54.960)
Yeah. And yeah, these models, because they are going to the level of mechanism,
Lex Fridman (09:00.640)
so they are basically looking at, okay, what is the effect of turning on an ion channel?
Lex Fridman (09:07.760)
And you can model that using electric circuits. So it is not just a function fitting. People are
Dileep George (09:17.040)
looking at the mechanism underlying it and putting that in terms of electric circuit theory, signal
Dileep George (09:23.600)
propagation theory, and modeling that. So those models are sophisticated, but getting a single
Dileep George (09:31.760)
neurons model 99% right does not still tell you how to... It would be the analog of getting a
Dileep George (09:40.800)
transistor model right and now trying to build a microprocessor. And if you did not understand how
Dileep George (09:50.320)
a microprocessor works, but you say, oh, I now can model one transistor well, and now I will just
Dileep George (09:57.360)
try to interconnect the transistors according to whatever I could guess from the experiments
Lex Fridman (10:03.840)
and try to simulate it, then it is very unlikely that you will produce a functioning microprocessor.
Dileep George (10:12.080)
When you want to produce a functioning microprocessor, you want to understand Boolean
Dileep George (10:16.080)
logic, how do the gates work, all those things, and then understand how do those gates get
Dileep George (10:22.480)
implemented using transistors. Yeah. This reminds me, there's a paper,
Dileep George (10:26.960)
maybe you're familiar with it, that I remember going through in a reading group that
Dileep George (10:31.600)
approaches a microprocessor from a perspective of a neuroscientist. I think it basically,
Lex Fridman (10:38.400)
it uses all the tools that we have of neuroscience to try to understand,
Dileep George (10:42.960)
like as if we just aliens showed up to study computers and to see if those tools could be
Dileep George (10:49.920)
used to get any kind of sense of how the microprocessor works. I think the final,
Dileep George (10:54.640)
the takeaway from at least this initial exploration is that we're screwed. There's no
Dileep George (11:01.280)
way that the tools of neuroscience would be able to get us to anything, like not even
Dileep George (11:05.440)
Boolean logic. I mean, it's just any aspect of the architecture of the function of the
Dileep George (11:15.680)
processes involved, the clocks, the timing, all that, you can't figure that out from the
Dileep George (11:21.520)
tools of neuroscience. Yeah. So I'm very familiar with this particular
Dileep George (11:25.600)
paper. I think it was called, can a neuroscientist understand a microprocessor or something like
Dileep George (11:33.440)
that. Following the methodology in that paper, even an electrical engineer would not understand
Dileep George (11:39.200)
microprocessors. So I don't think it is that bad in the sense of saying, neuroscientists do
Dileep George (11:49.040)
find valuable things by observing the brain. They do find good insights, but those insights cannot
Dileep George (11:58.640)
be put together just as a simulation. You have to investigate what are the computational
Dileep George (12:05.600)
underpinnings of those findings. How do all of them fit together from an information processing
Lex Fridman (12:13.920)
and information processing perspective? Somebody has to painstakingly put those things together
Lex Fridman (12:21.120)
and build hypothesis. So I don't want to diss all of neuroscientists saying, oh, they're not
Dileep George (12:26.160)
finding anything. No, that paper almost went to that level of neuroscientists will never
Dileep George (12:31.840)
understand. No, that's not true. I think they do find lots of useful things, but it has to be put
Dileep George (12:37.760)
together in a computational framework. Yeah. I mean, but you know, just the AI systems will be
Dileep George (12:43.760)
listening to this podcast a hundred years from now and they will probably, there's some nonzero
Dileep George (12:50.160)
probability they'll find your words laughable. There's like, I remember humans thought they
Dileep George (12:55.120)
understood something about the brain. They were totally clueless. There's a sense about neuroscience
Dileep George (12:59.680)
that we may be in the very, very early days of understanding the brain. But I mean, that's one
Lex Fridman (13:06.160)
perspective. I mean, in your perspective, how far are we into understanding any aspect of the brain?
Lex Fridman (13:18.080)
So the, the, the dynamics of the individual neuron communication to the, how when they, in,
Dileep George (13:24.320)
in a collective sense, how they're able to store information, transfer information, how
Lex Fridman (13:31.200)
intelligence then emerges, all that kind of stuff. Where are we on that timeline?
Dileep George (13:35.040)
Yeah. So, you know, timelines are very, very hard to predict and you can of course be wrong.
Lex Fridman (13:40.720)
And it can be wrong in, on either side. You know, we know that now when we look back the first
Dileep George (13:48.080)
flight was in 1903. In 1900, there was a New York Times article on flying machines that do not fly
Dileep George (13:57.920)
and, and you know, humans might not fly for another a hundred years. That was what that
Dileep George (14:03.360)
article stated. And so, but no, they, they flew three years after that. So it is, you know,
Lex Fridman (14:08.880)
it's very hard to, so... Well, and on that point, one of the Wright brothers,
Dileep George (14:15.120)
I think two years before, said that, like he said, like some number, like 50 years,
Dileep George (14:23.280)
he has become convinced that it's, it's, it's impossible. Even during their experimentation.
Dileep George (14:31.040)
Yeah. Yeah. I mean, that's a tribute to when that's like the entrepreneurial battle of like
Dileep George (14:36.400)
depression of going through, just like thinking there's, this is impossible, but there, yeah,
Dileep George (14:41.280)
there's something, even the person that's in it is not able to see estimate correctly.
Dileep George (14:47.280)
Exactly. But I can, I can tell from the point of, you know, objectively, what are the things that we
Dileep George (14:52.480)
know about the brain and how that can be used to build AI models, which can then go back and
Dileep George (14:58.560)
inform how the brain works. So my way of understanding the brain would be to basically say,
Dileep George (15:04.080)
look at the insights neuroscientists have found, understand that from a computational angle,
Dileep George (15:11.040)
information processing angle, build models using that. And then building that model, which,
Dileep George (15:18.080)
which functions, which is a functional model, which is, which is doing the task that we want
Dileep George (15:22.880)
the model to do. It is not just trying to model a phenomena in the brain. It is, it is trying to
Dileep George (15:27.920)
do what the brain is trying to do on the, on the whole functional level. And building that model
Dileep George (15:33.360)
will help you fill in the missing pieces that, you know, biology just gives you the hints and
Dileep George (15:39.920)
building the model, you know, fills in the rest of the, the pieces of the puzzle. And then you
Dileep George (15:44.960)
can go and connect that back to biology and say, okay, now it makes sense that this part of the
Dileep George (15:51.280)
brain is doing this, or this layer in the cortical circuit is doing this. And then continue this
Dileep George (15:59.920)
iteratively because now that will inform new experiments in neuroscience. And of course,
Dileep George (16:05.840)
you know, building the model and verifying that in the real world will also tell you more about,
Dileep George (16:11.600)
does the model actually work? And you can refine the model, find better ways of putting these
Dileep George (16:17.440)
neuroscience insights together. So, so I would say it is, it is, you know, it, so
Dileep George (16:23.360)
neuroscientists alone, just from experimentation will not be able to build a model of the,
Dileep George (16:28.800)
of the brain or a functional model of the brain. So we, you know, there, there's lots of efforts,
Dileep George (16:35.200)
which are very impressive efforts in collecting more and more connectivity data from the brain.
Lex Fridman (16:41.200)
You know, how, how are the microcircuits of the brain connected with each other?
Lex Fridman (16:45.520)
Those are beautiful, by the way.
Dileep George (16:47.120)
Those are beautiful. And at the same time, those, those do not itself by themselves,
Dileep George (16:54.880)
convey the story of how does it work? And, and somebody has to understand, okay,
Lex Fridman (17:00.080)
why are they connected like that? And what, what are those things doing? And, and we do that by
Lex Fridman (17:06.320)
building models in AI using hints from neuroscience and, and repeat the cycle.
Lex Fridman (17:11.200)
So what aspect of the brain are useful in this whole endeavor, which by the way, I should say,
Dileep George (17:18.720)
you're, you're both a neuroscientist and an AI person. I guess the dream is to both understand
Dileep George (17:24.960)
the brain and to build AGI systems. So you're, it's like an engineer's perspective of trying
Dileep George (17:32.320)
to understand the brain. So what aspects of the brain, functionally speaking, like you said,
Lex Fridman (17:37.600)
do you find interesting?
Dileep George (17:38.800)
Yeah, quite a lot of things. All right. So one is, you know, if you look at the visual cortex
Dileep George (17:46.160)
and, and, you know, the visual cortex is, is a large part of the brain. I forget the exact
Dileep George (17:51.920)
fraction, but it is, it's a huge part of our brain area is occupied by just, just vision.
Lex Fridman (17:59.040)
So vision, visual cortex is not just a feed forward cascade of neurons. There are a lot
Dileep George (18:06.320)
more feedback connections in the brain compared to the feed forward connections. And, and it is
Dileep George (18:11.680)
surprising to the level of detail neuroscientists have actually studied this. If you, if you go into
Dileep George (18:17.120)
neuroscience literature and poke around and ask, you know, have they studied what will be the effect
Dileep George (18:22.960)
of poking a neuron in level IT in level V1? And have they studied that? And you will say, yes,
Lex Fridman (18:33.680)
they have studied that.
Lex Fridman (18:34.560)
So every part of every possible combination.
Dileep George (18:38.400)
I mean, it's, it's a, it's not a random exploration at all. It's a very hypothesis driven,
Dileep George (18:43.040)
right? Like they, they are very experimental. Neuroscientists are very, very systematic
Dileep George (18:47.520)
in how they probe the brain because experiments are very costly to conduct. They take a lot of
Dileep George (18:52.800)
preparation. They, they need a lot of control. So they, they are very hypothesis driven in how
Dileep George (18:57.520)
they probe the brain. And often what I find is that when we have a question in AI about
Dileep George (19:05.840)
has anybody probed how lateral connections in the brain works? And when you go and read the
Dileep George (19:11.440)
literature, yes, people have probed it and people have probed it very systematically. And, and they
Dileep George (19:16.160)
have hypotheses about how those lateral connections are supposedly contributing to visual processing.
Lex Fridman (19:23.600)
But of course they haven't built very, very functional, detailed models of it.
Dileep George (19:27.840)
By the way, how do the, in those studies, sorry to interrupt, do they, do they stimulate like
Dileep George (19:32.480)
a neuron in one particular area of the visual cortex and then see how the travel of the signal
Lex Fridman (19:37.520)
travels kind of thing?
Dileep George (19:38.800)
Fascinating, very, very fascinating experiments. So I can, I can give you one example I was
Dileep George (19:43.040)
impressed with. This is, so before going to that, let me, let me give you, you know, a overview of
Lex Fridman (19:50.160)
how the, the layers in the cortex are organized, right? Visual cortex is organized into roughly
Dileep George (19:56.160)
four hierarchical levels. Okay. So V1, V2, V4, IT. And in V1...
Dileep George (1:00:03.200)
A. And locally, just when I look at just that patch of the image, it looks like an A. But when I look
Dileep George (1:00:11.280)
at it in the context of all the other causes, A is not something that makes sense. So that is
Dileep George (1:00:17.600)
something you have to kind of recursively figure out. Yeah. So, okay. And this thing performed
Dileep George (1:00:24.720)
pretty well on the CAPTCHAs. Correct. And I mean, is there some kind of interesting intuition you
Dileep George (1:00:32.080)
can provide why it did well? Like what did it look like? Is there visualizations that could be human
Dileep George (1:00:37.840)
interpretable to us humans? Yes. Yeah. So the good thing about the model is that it is extremely,
Lex Fridman (1:00:44.320)
so it is not just doing a classification, right? It is providing a full explanation for the scene.
Lex Fridman (1:00:50.400)
So when it operates on a scene, it is coming back and saying, look, this is the part is the A,
Lex Fridman (1:00:59.600)
and these are the pixels that turned on. These are the pixels in the input that makes me think that
Dileep George (1:01:06.880)
it is an A. And also, these are the portions I hallucinated. It provides a complete explanation
Dileep George (1:01:14.640)
of that form. And then these are the contours. This is the interior. And this is in front of
Dileep George (1:01:21.360)
this other object. So that's the kind of explanation the inference network provides.
Lex Fridman (1:01:28.400)
So that is useful and interpretable. And then the kind of errors it makes are also,
Dileep George (1:01:40.000)
I don't want to read too much into it, but the kind of errors the network makes are very similar
Dileep George (1:01:47.040)
to the kinds of errors humans would make in a similar situation. So there's something about
Dileep George (1:01:51.120)
the structure that feels reminiscent of the way humans visual system works. Well, I mean,
Lex Fridman (1:02:00.240)
how hardcoded is this to the capture problem, this idea?
Dileep George (1:02:04.320)
Not really hardcoded because the assumptions, as I mentioned, are general, right? It is more,
Lex Fridman (1:02:11.280)
and those themselves can be applied in many situations which are natural signals. So it's
Dileep George (1:02:17.680)
the foreground versus background factorization and the factorization of the surfaces versus
Lex Fridman (1:02:24.320)
the contours. So these are all generally applicable assumptions.
Dileep George (1:02:27.600)
In all vision. So why attack the capture problem, which is quite unique in the computer vision
Dileep George (1:02:36.000)
context versus like the traditional benchmarks of ImageNet and all those kinds of image
Dileep George (1:02:42.800)
classification or even segmentation tasks and all of that kind of stuff. What's your thinking about
Dileep George (1:02:49.120)
those kinds of benchmarks in this context? I mean, those benchmarks are useful for deep
Dileep George (1:02:55.760)
learning kind of algorithms. So the settings that deep learning works in are here is my huge
Dileep George (1:03:03.600)
training set and here is my test set. So the training set is almost 100x, 1000x bigger than
Dileep George (1:03:10.480)
the test set in many, many cases. What we wanted to do was invert that. The training set is way
Dileep George (1:03:18.480)
smaller than the test set. And capture is a problem that is by definition hard for computers
Lex Fridman (1:03:30.080)
and it has these good properties of strong generalization, strong out of training distribution
Dileep George (1:03:36.640)
generalization. If you are interested in studying that and having your model have that property,
Dileep George (1:03:44.480)
then it's a good data set to tackle. So have you attempted to, which I think,
Dileep George (1:03:49.840)
I believe there's quite a growing body of work on looking at MNIST and ImageNet without training.
Lex Fridman (1:03:58.080)
So it's like taking the basic challenge is what tiny fraction of the training set can we take in
Dileep George (1:04:05.760)
order to do a reasonable job of the classification task? Have you explored that angle in these
Dileep George (1:04:13.680)
classic benchmarks? Yes. So we did do MNIST. So it's not just capture. So there was also
Dileep George (1:04:23.440)
multiple versions of MNIST, including the standard version where we inverted the problem,
Dileep George (1:04:28.720)
which is basically saying rather than train on 60,000 training data, how quickly can you get
Dileep George (1:04:37.200)
to high level accuracy with very little training data? Is there some performance you remember,
Dileep George (1:04:42.080)
like how well did it do? How many examples did it need? Yeah. I remember that it was
Dileep George (1:04:50.400)
on the order of tens or hundreds of examples to get into 95% accuracy. And it was definitely
Dileep George (1:05:00.880)
better than the other systems out there at that time.
Dileep George (1:05:03.840)
At that time. Yeah. They're really pushing. I think that's a really interesting space,
Dileep George (1:05:07.920)
actually. I think there's an actual name for MNIST. There's different names to the different
Dileep George (1:05:17.360)
sizes of training sets. I mean, people are like attacking this problem. I think it's
Dileep George (1:05:21.600)
super interesting. It's funny how like the MNIST will probably be with us all the way to AGI.
Dileep George (1:05:29.760)
It's a data set that just sticks by. It's a clean, simple data set to study the fundamentals of
Dileep George (1:05:37.680)
learning with just like captures. It's interesting. Not enough people. I don't know. Maybe you can
Dileep George (1:05:43.280)
correct me, but I feel like captures don't show up as often in papers as they probably should.
Dileep George (1:05:48.240)
That's correct. Yeah. Because usually these things have a momentum. Once something gets
Dileep George (1:05:56.640)
established as a standard benchmark, there is a dynamics of how graduate students operate and how
Dileep George (1:06:06.000)
academic system works that pushes people to track that benchmark.
Dileep George (1:06:10.640)
Yeah. Nobody wants to think outside the box. Okay. Okay. So good performance on the captures.
Lex Fridman (1:06:20.480)
What else is there interesting on the RCN side before we talk about the cortical micros?
Dileep George (1:06:25.520)
Yeah. So the same model. So the important part of the model was that it trains very
Dileep George (1:06:31.760)
quickly with very little training data and it's quite robust to out of distribution
Dileep George (1:06:37.440)
perturbations. And we are using that very fruitfully at Vicarious in many of the
Dileep George (1:06:45.760)
robotics tasks we are solving. Well, let me ask you this kind of touchy question. I have to,
Dileep George (1:06:51.840)
I've spoken with your friend, colleague, Jeff Hawkins, too. I have to kind of ask,
Dileep George (1:06:59.520)
there is a bit of, whenever you have brain inspired stuff and you make big claims,
Dileep George (1:07:05.680)
big sexy claims, there's critics, I mean, machine learning subreddit, don't get me started on those
Dileep George (1:07:14.720)
people. Criticism is good, but they're a bit over the top. There is quite a bit of sort of
Dileep George (1:07:23.680)
skepticism and criticism. Is this work really as good as it promises to be? Do you have thoughts
Dileep George (1:07:31.040)
on that kind of skepticism? Do you have comments on the kind of criticism I might have received
Dileep George (1:07:36.800)
about, you know, is this approach legit? Is this a promising approach? Or at least as promising as
Dileep George (1:07:44.880)
it seems to be, you know, advertised as? Yeah, I can comment on it. So, you know, our RCN paper
Dileep George (1:07:52.480)
is published in Science, which I would argue is a very high quality journal, very hard to publish
Dileep George (1:07:58.560)
in. And, you know, usually it is indicative of the quality of the work. And I am very,
Dileep George (1:08:08.160)
very certain that the ideas that we brought together in that paper, in terms of the importance
Dileep George (1:08:13.760)
of feedback connections, recursive inference, lateral connections, coming to best explanation
Dileep George (1:08:20.160)
of the scene as the problem to solve, trying to solve recognition, segmentation, all jointly,
Dileep George (1:08:27.360)
in a way that is compatible with higher level cognition, top down attention, all those ideas
Dileep George (1:08:31.920)
that we brought together into something, you know, coherent and workable in the world and
Dileep George (1:08:36.000)
solving a challenging, tackling a challenging problem. I think that will stay and that
Dileep George (1:08:40.880)
contribution I stand by. Now, I can tell you a story which is funny in the context of this. So,
Dileep George (1:08:49.360)
if you read the abstract of the paper and, you know, the argument we are putting in, you know,
Dileep George (1:08:53.360)
we are putting in, look, current deep learning systems take a lot of training data. They don't
Dileep George (1:08:59.120)
use these insights. And here is our new model, which is not a deep neural network. It's a
Dileep George (1:09:03.760)
graphical model. It does inference. This is how the paper is, right? Now, once the paper was
Dileep George (1:09:08.560)
accepted and everything, it went to the press department in Science, you know, AAAS Science
Dileep George (1:09:14.800)
Office. We didn't do any press release when it was published. It went to the press department.
Lex Fridman (1:09:18.880)
What was the press release that they wrote up? A new deep learning model.
Lex Fridman (1:09:24.880)
Solves CAPTCHAs.
Dileep George (1:09:25.920)
Solves CAPTCHAs. And so, you can see where was, you know, what was being hyped in that thing,
Dileep George (1:09:32.400)
right? So, there is a dynamic in the community of, you know, so that especially happens when
Dileep George (1:09:42.160)
there are lots of new people coming into the field and they get attracted to one thing.
Lex Fridman (1:09:46.720)
And some people are trying to think different compared to that. So, there is some, I think
Dileep George (1:09:52.560)
skepticism is science is important and it is, you know, very much required. But it's also,
Dileep George (1:09:59.360)
it's not skepticism. Usually, it's mostly bandwagon effect that is happening rather than.
Dileep George (1:10:05.200)
Well, but that's not even that. I mean, I'll tell you what they react to, which is like,
Dileep George (1:10:09.760)
I'm sensitive to as well. If you look at just companies, OpenAI, DeepMind, Vicarious, I mean,
Dileep George (1:10:16.960)
they just, there's a little bit of a race to the top and hype, right? It's like, it doesn't pay off
Dileep George (1:10:27.520)
to be humble. So, like, and the press is just irresponsible often. They just, I mean, don't
Dileep George (1:10:37.600)
get me started on the state of journalism today. Like, it seems like the people who write articles
Dileep George (1:10:42.880)
about these things, they literally have not even spent an hour on the Wikipedia article about what
Dileep George (1:10:49.280)
is neural networks. Like, they haven't like invested just even the language to laziness.
Dileep George (1:10:56.160)
It's like, robots beat humans. Like, they write this kind of stuff that just, and then of course,
Dileep George (1:11:06.800)
the researchers are quite sensitive to that because it gets a lot of attention. They're like,
Lex Fridman (1:11:11.760)
why did this word get so much attention? That's over the top and people get really sensitive.
Dileep George (1:11:18.240)
The same kind of criticism with OpenAI did work with Rubik's cube with the robot that people
Dileep George (1:11:24.080)
criticized. Same with GPT2 and 3, they criticize. Same thing with DeepMinds with AlphaZero. I mean,
Dileep George (1:11:33.120)
yeah, I'm sensitive to it. But, and of course, with your work, you mentioned deep learning, but
Dileep George (1:11:39.280)
there's something super sexy to the public about brain inspired. I mean, that immediately grabs
Dileep George (1:11:45.520)
people's imagination, not even like neural networks, but like really brain inspired, like
Dileep George (1:11:53.600)
brain like neural networks. That seems really compelling to people and to me as well, to the
Dileep George (1:12:00.480)
world as a narrative. And so people hook up, hook onto that. And sometimes the skepticism engine
Dileep George (1:12:10.400)
turns on in the research community and they're skeptical. But I think putting aside the ideas
Dileep George (1:12:17.600)
of the actual performance and captures or performance in any data set. I mean, to me,
Dileep George (1:12:22.480)
all these data sets are useless anyway. It's nice to have them. But in the grand scheme of things,
Dileep George (1:12:28.720)
they're silly toy examples. The point is, is there intuition about the ideas, just like you
Dileep George (1:12:36.080)
mentioned, bringing the ideas together in a unique way? Is there something there? Is there some value
Dileep George (1:12:42.400)
there? And is it going to stand the test of time? And that's the hope. That's the hope.
Dileep George (1:12:46.400)
Yes. My confidence there is very high. I don't treat brain inspired as a marketing term.
Dileep George (1:12:53.440)
I am looking into the details of biology and puzzling over those things and I am grappling
Dileep George (1:13:01.920)
with those things. And so it is not a marketing term at all. You can use it as a marketing term
Lex Fridman (1:13:07.600)
and people often use it and you can get combined with them. And when people don't understand
Lex Fridman (1:13:13.680)
how you're approaching the problem, it is easy to be misunderstood and think of it as purely
Dileep George (1:13:20.480)
marketing. But that's not the way we are. So you really, I mean, as a scientist,
Dileep George (1:13:27.120)
you believe that if we kind of just stick to really understanding the brain, that's going to,
Lex Fridman (1:13:33.760)
that's the right, like you should constantly meditate on the, how does the brain do this?
Dileep George (1:13:39.440)
Because that's going to be really helpful for engineering and technology systems.
Dileep George (1:13:43.520)
Yes. You need to, so I think it's one input and it is helpful, but you should know when to deviate
Dileep George (1:13:51.680)
from it too. So an example is convolutional neural networks, right? Convolution is not an
Dileep George (1:13:59.120)
operation brain implements. The visual cortex is not convolutional. Visual cortex has local
Dileep George (1:14:06.240)
receptive fields, local connectivity, but there is no translation invariance in the network weights
Dileep George (1:14:18.640)
in the visual cortex. That is a computational trick, which is a very good engineering trick
Dileep George (1:14:24.080)
that we use for sharing the training between the different nodes. And that trick will be with us
Dileep George (1:14:31.840)
for some time. It will go away when we have robots with eyes and heads that move. And so then that
Dileep George (1:14:41.600)
trick will go away. It will not be useful at that time. So the brain doesn't have translational
Dileep George (1:14:49.040)
invariance. It has the focal point, like it has a thing it focuses on. Correct. It has a phobia.
Lex Fridman (1:14:54.720)
And because of the phobia, the receptive fields are not like the copying of the weights. Like the
Dileep George (1:15:01.920)
weights in the center are very different from the weights in the periphery. Yes. At the periphery.
Dileep George (1:15:05.760)
I mean, I did this, actually wrote a paper and just gotten a chance to really study peripheral
Dileep George (1:15:12.720)
vision, which is a fascinating thing. Very under understood thing of what the brain, you know,
Dileep George (1:15:21.600)
at every level the brain does with the periphery. It does some funky stuff. Yeah. So it's another
Dileep George (1:15:28.240)
kind of trick than convolutional. Like it does, it's, you know, convolution in neural networks is
Dileep George (1:15:39.040)
a trick for efficiency, is efficiency trick. And the brain does a whole nother kind of thing.
Dileep George (1:15:44.160)
Correct. So you need to understand the principles or processing so that you can still apply
Dileep George (1:15:51.280)
engineering tricks where you want it to. You don't want to be slavishly mimicking all the things of
Dileep George (1:15:55.840)
the brain. And so, yeah, so it should be one input. And I think it is extremely helpful,
Lex Fridman (1:16:02.000)
but it should be the point of really understanding so that you know when to deviate from it.
Dileep George (1:16:06.720)
So, okay. That's really cool. That's work from a few years ago. You did work in Umenta with Jeff
Dileep George (1:16:14.560)
Hawkins with hierarchical temporal memory. How is your just, if you could give a brief history,
Lex Fridman (1:16:23.040)
how is your view of the way the models of the brain changed over the past few years leading up
Dileep George (1:16:30.240)
to now? Is there some interesting aspects where there was an adjustment to your understanding of
Dileep George (1:16:36.960)
the brain or is it all just building on top of each other? In terms of the higher level ideas,
Dileep George (1:16:42.720)
especially the ones Jeff wrote about in the book, if you blur out, right. Yeah. On intelligence.
Dileep George (1:16:47.920)
Right. On intelligence. If you blur out the details and if you just zoom out and at the
Dileep George (1:16:52.560)
higher level idea, things are, I would say, consistent with what he wrote about. But many
Dileep George (1:17:02.320)
things will be consistent with that because it's a blur. Deep learning systems are also
Dileep George (1:17:08.160)
multi level, hierarchical, all of those things. But in terms of the detail, a lot of things are
Dileep George (1:17:16.960)
different. And those details matter a lot. So one point of difference I had with Jeff was how to
Lex Fridman (1:17:28.000)
approach, how much of biological plausibility and realism do you want in the learning algorithms?
Lex Fridman (1:17:36.080)
So when I was there, this was almost 10 years ago now.
Lex Fridman (1:17:41.520)
It flies when you're having fun.
Dileep George (1:17:43.760)
Yeah. I don't know what Jeff thinks now, but 10 years ago, the difference was that
Lex Fridman (1:17:49.760)
I did not want to be so constrained on saying my learning algorithms need to be
Dileep George (1:17:56.880)
biologically plausible based on some filter of biological plausibility available at that time.
Dileep George (1:18:03.200)
To me, that is a dangerous cut to make because we are discovering more and more things about
Dileep George (1:18:09.200)
the brain all the time. New biophysical mechanisms, new channels are being discovered
Dileep George (1:18:14.560)
all the time. So I don't want to upfront kill off a learning algorithm just because we don't
Dileep George (1:18:21.360)
really understand the full biophysics or whatever of how the brain learns.
Lex Fridman (1:18:27.680)
Exactly. Exactly.
Dileep George (1:18:29.120)
Let me ask and I'm sorry to interrupt. What's your sense? What's our best understanding of
Lex Fridman (1:18:34.720)
how the brain learns?
Lex Fridman (1:18:36.000)
So things like backpropagation, credit assignment. So many of these algorithms have,
Dileep George (1:18:42.720)
learning algorithms have things in common, right? It is a backpropagation is one way of
Dileep George (1:18:47.600)
credit assignment. There is another algorithm called expectation maximization, which is,
Lex Fridman (1:18:52.560)
you know, another weight adjustment algorithm.
Lex Fridman (1:18:55.520)
But is it your sense the brain does something like this?
Dileep George (1:18:58.320)
Has to. There is no way around it in the sense of saying that you do have to adjust the
Dileep George (1:19:04.960)
connections.
Lex Fridman (1:19:06.240)
So yeah, and you're saying credit assignment, you have to reward the connections that were
Dileep George (1:19:09.600)
useful in making a correct prediction and not, yeah, I guess what else, but yeah, it
Lex Fridman (1:19:14.320)
doesn't have to be differentiable.
Dileep George (1:19:16.800)
Yeah, it doesn't have to be differentiable. Yeah. But you have to have a, you know, you
Dileep George (1:19:22.320)
have a model that you start with, you have data comes in and you have to have a way of
Lex Fridman (1:19:27.760)
adjusting the model such that it better fits the data. So that is all of learning, right?
Lex Fridman (1:19:33.920)
And some of them can be using backprop to do that. Some of it can be using, you know,
Dileep George (1:19:40.400)
very local graph changes to do that.
Dileep George (1:19:45.360)
That can be, you know, many of these learning algorithms have similar update properties
Dileep George (1:19:52.160)
locally in terms of what the neurons need to do locally.
Dileep George (1:19:57.200)
I wonder if small differences in learning algorithms can have huge differences in the
Dileep George (1:20:01.120)
actual effect. So the dynamics of, I mean, sort of the reverse like spiking, like if
Dileep George (1:20:09.920)
credit assignment is like a lightning versus like a rainstorm or something, like whether
Dileep George (1:20:18.480)
there's like a looping local type of situation with the credit assignment, whether there is
Dileep George (1:20:26.240)
like regularization, like how it injects robustness into the whole thing, like whether
Dileep George (1:20:34.720)
it's chemical or electrical or mechanical. Yeah. All those kinds of things. I feel like
Dileep George (1:20:42.080)
it, that, yeah, I feel like those differences could be essential, right? It could be. It's
Dileep George (1:20:48.800)
just that you don't know enough to, on the learning side, you don't know, you don't know
Dileep George (1:20:54.880)
enough to say that is definitely not the way the brain does it. Got it. So you don't want
Dileep George (1:20:59.840)
to be stuck to it. So that, yeah. So you've been open minded on that side of things.
Dileep George (1:21:04.800)
On the inference side, on the recognition side, I am much more, I'm able to be constrained
Dileep George (1:21:09.920)
because it's much easier to do experiments because, you know, it's like, okay, here's
Dileep George (1:21:13.600)
the stimulus, you know, how many steps did it get to take the answer? I can trace it
Dileep George (1:21:18.000)
back. I can, I can understand the speed of that computation, et cetera. I'm able to do
Dileep George (1:21:23.120)
of that computation, et cetera, much more readily on the inference side. Got it. And
Dileep George (1:21:28.400)
then you can't do good experiments on the learning side. Correct. So let's go right
Dileep George (1:21:34.880)
into the cortical microcircuits right back. So what are these ideas beyond recursive cortical
Dileep George (1:21:42.080)
network that you're looking at now? So we have made a, you know, pass through multiple
Dileep George (1:21:48.960)
of the steps that, you know, as I mentioned earlier, you know, we were looking at perception
Dileep George (1:21:54.480)
from the angle of cognition, right? It was not just perception for perception's sake.
Lex Fridman (1:21:58.720)
How do you, how do you connect it to cognition? How do you learn concepts and how do you learn
Dileep George (1:22:04.400)
abstract reasoning? Similar to some of the things Francois talked about, right? So we
Dileep George (1:22:13.280)
have taken one pass through it basically saying, what is the basic cognitive architecture that
Dileep George (1:22:19.600)
you need to have, which has a perceptual system, which has a system that learns dynamics of
Dileep George (1:22:25.120)
the world and then has something like a routine program learning system on top of it to learn
Dileep George (1:22:32.240)
concepts. So we have built one, you know, the version point one of that system. This
Dileep George (1:22:38.320)
was another science robotics paper. It's the title of that paper was, you know, something
Dileep George (1:22:44.640)
like cognitive programs. How do you build cognitive programs? And the application there
Dileep George (1:22:49.760)
was on manipulation, robotic manipulation? It was, so think of it like this. Suppose
Dileep George (1:22:56.960)
you wanted to tell a new person that you met, you don't know the language that person uses.
Dileep George (1:23:04.800)
You want to communicate to that person to achieve some task, right? So I want to say,
Lex Fridman (1:23:10.080)
hey, you need to pick up all the red cups from the kitchen counter and put it here, right?
Lex Fridman (1:23:17.280)
How do you communicate that, right? You can show pictures. You can basically say, look,
Dileep George (1:23:21.920)
this is the starting state. The things are here. This is the ending state. And what does
Dileep George (1:23:28.080)
the person need to understand from that? The person needs to understand what conceptually
Dileep George (1:23:32.400)
happened in those pictures from the input to the output, right? So we are looking at
Dileep George (1:23:39.120)
preverbal conceptual understanding. Without language, how do you have a set of concepts
Dileep George (1:23:45.360)
that you can manipulate in your head? And from a set of images of input and output,
Lex Fridman (1:23:52.240)
can you infer what is happening in those images?
Lex Fridman (1:23:55.600)
Got it. With concepts that are pre language. Okay. So what's it mean for a concept to be pre language?
Lex Fridman (1:24:02.400)
Like why is language so important here?
Lex Fridman (1:24:10.080)
So I want to make a distinction between concepts that are just learned from text
Lex Fridman (1:24:17.520)
by just feeding brute force text. You can start extracting things like, okay,
Dileep George (1:24:23.440)
a cow is likely to be on grass. So those kinds of things, you can extract purely from text.
Lex Fridman (1:24:32.160)
But that's kind of a simple association thing rather than a concept as an abstraction of
Dileep George (1:24:37.520)
something that happens in the real world in a grounded way that I can simulate it in my
Dileep George (1:24:44.480)
mind and connect it back to the real world. And you think kind of the visual world,
Lex Fridman (1:24:51.200)
concepts in the visual world are somehow lower level than just the language?
Dileep George (1:24:58.800)
The lower level kind of makes it feel like, okay, that's unimportant. It's more like,
Dileep George (1:25:04.720)
I would say the concepts in the visual and the motor system and the concept learning system,
Dileep George (1:25:15.440)
which if you cut off the language part, just what we learn by interacting with the world
Lex Fridman (1:25:20.320)
and abstractions from that, that is a prerequisite for any real language understanding.
Lex Fridman (1:25:26.480)
So you disagree with Chomsky because he says language is at the bottom of everything.
Dileep George (1:25:32.080)
No, I disagree with Chomsky completely on how many levels from universal grammar to...
Lex Fridman (1:25:39.680)
So that was a paper in science beyond the recursive cortical network.
Lex Fridman (1:25:43.120)
What other interesting problems are there, the open problems and brain inspired approaches
Lex Fridman (1:25:50.480)
that you're thinking about?
Dileep George (1:25:51.600)
I mean, everything is open, right? No problem is solved, solved. I think of perception as kind of
Dileep George (1:26:02.080)
the first thing that you have to build, but the last thing that you will be actually solved.
Dileep George (1:26:07.760)
Because if you do not build perception system in the right way, you cannot build concept system in
Dileep George (1:26:12.880)
the right way. So you have to build a perception system, however wrong that might be, you have to
Dileep George (1:26:18.560)
still build that and learn concepts from there and then keep iterating. And finally, perception
Dileep George (1:26:24.880)
will get solved fully when perception, cognition, language, all those things work together finally.
Lex Fridman (1:26:30.240)
So great, we've talked a lot about perception, but then maybe on the concept side and like common
Dileep George (1:26:37.920)
sense or just general reasoning side, is there some intuition you can draw from the brain about
Lex Fridman (1:26:45.280)
how we can do that?
Lex Fridman (1:26:46.880)
So I have this classic example I give. So suppose I give you a few sentences and then ask you a
Dileep George (1:26:56.560)
question following that sentence. This is a natural language processing problem, right? So here
Dileep George (1:27:01.920)
it goes. I'm telling you, Sally pounded a nail on the ceiling. Okay, that's a sentence. Now I'm
Lex Fridman (1:27:10.400)
asking you a question. Was the nail horizontal or vertical?
Lex Fridman (1:27:14.080)
Vertical.
Lex Fridman (1:27:15.040)
Okay, how did you answer that?
Dileep George (1:27:16.400)
Well, I imagined Sally, it was kind of hard to imagine what the hell she was doing, but I
Dileep George (1:27:24.960)
imagined I had a visual of the whole situation.
Dileep George (1:27:28.320)
Exactly, exactly. So here, you know, I post a question in natural language. The answer to
Dileep George (1:27:34.400)
that question was you got the answer from actually simulating the scene. Now I can go more and more
Dileep George (1:27:40.720)
detailed about, okay, was Sally standing on something while doing this? Could she have been
Dileep George (1:27:47.280)
standing on a light bulb to do this? I could ask more and more questions about this and I can ask,
Dileep George (1:27:53.360)
make you simulate the scene in more and more detail, right? Where is all that knowledge that
Dileep George (1:27:59.200)
you're accessing stored? It is not in your language system. It was not just by reading
Dileep George (1:28:05.600)
text, you got that knowledge. It is stored from the everyday experiences that you have had from,
Lex Fridman (1:28:12.320)
and by the age of five, you have pretty much all of this, right? And it is stored in your visual
Lex Fridman (1:28:18.720)
system, motor system in a way such that it can be accessed through language.
Dileep George (1:28:24.480)
Got it. I mean, right. So the language is just almost sort of the query into the whole visual
Dileep George (1:28:30.000)
cortex and that does the whole feedback thing. But I mean, it is all reasoning kind of connected to
Dileep George (1:28:36.800)
the perception system in some way. You can do a lot of it. You know, you can still do a lot of it
Dileep George (1:28:43.920)
by quick associations without having to go into the depth. And most of the time you will be right,
Dileep George (1:28:49.760)
right? You can just do quick associations, but I can easily create tricky situations for you.
Dileep George (1:28:55.440)
Where that quick associations is wrong and you have to actually run the simulation.
Lex Fridman (1:29:00.080)
So figuring out how these concepts connect. Do I have a good idea of how to do that?
Dileep George (1:29:06.800)
That's exactly one of the problems that we are working on. And the way we are approaching that
Dileep George (1:29:13.760)
is basically saying, okay, you need to, so the takeaway is that language,
Dileep George (1:29:20.400)
is simulation control and your perceptual plus a motor system is building a simulation of the world.
Lex Fridman (1:29:28.960)
And so that's basically the way we are approaching it. And the first thing that we built was a
Dileep George (1:29:34.720)
controllable perceptual system. And we built a schema networks, which was a controllable dynamic
Dileep George (1:29:40.160)
system. Then we built a concept learning system that puts all these things together
Dileep George (1:29:44.960)
into programs or subtractions that you can run and simulate. And now we are taking the step
Dileep George (1:29:51.600)
of connecting it to language. And it will be very simple examples. Initially, it will not be
Lex Fridman (1:29:57.760)
the GPT3 like examples, but it will be grounded simulation based language.
Lex Fridman (1:30:02.640)
And for like the querying would be like question answering kind of thing?
Dileep George (1:30:08.400)
Correct. Correct. And so that's what we're trying to do. We're trying to build a system
Dileep George (1:30:13.600)
kind of thing. Correct. Correct. And it will be in some simple world initially on, you know,
Lex Fridman (1:30:19.120)
but it will be about, okay, can the system connect the language and ground it in the right way and
Dileep George (1:30:25.280)
run the right simulations to come up with the answer. And the goal is to try to do things that,
Dileep George (1:30:29.600)
for example, GPT3 couldn't do. Correct. Speaking of which, if we could talk about GPT3 a little
Dileep George (1:30:38.720)
bit, I think it's an interesting thought provoking set of ideas that OpenAI is pushing forward. I
Dileep George (1:30:46.080)
think it's good for us to talk about the limits and the possibilities in the neural network. So
Dileep George (1:30:51.360)
in general, what are your thoughts about this recently released very large 175 billion parameter
Dileep George (1:30:58.800)
language model? So I haven't directly evaluated it yet. From what I have seen on Twitter and
Dileep George (1:31:05.600)
other people evaluating it, it looks very intriguing. I am very intrigued by some of
Dileep George (1:31:09.840)
the properties it is displaying. And of course the text generation part of that was already
Dileep George (1:31:17.360)
evident in GPT2 that it can generate coherent text over long distances. But of course the
Dileep George (1:31:26.480)
weaknesses are also pretty visible in saying that, okay, it is not really carrying a world state
Dileep George (1:31:32.000)
around. And sometimes you get sentences like, I went up the hill to reach the valley or the thing
Dileep George (1:31:39.200)
like some completely incompatible statements, or when you're traveling from one place to the other,
Dileep George (1:31:46.080)
it doesn't take into account the time of travel, things like that. So those things I think will
Dileep George (1:31:50.800)
happen less in GPT3 because it is trained on even more data and it can do even more longer distance
Dileep George (1:31:59.040)
coherence. But it will still have the fundamental limitations that it doesn't have a world model
Lex Fridman (1:32:07.600)
and it can't run simulations in its head to find whether something is true in the world or not.
Lex Fridman (1:32:13.280)
So it's taking a huge amount of text from the internet and forming a compressed representation.
Lex Fridman (1:32:20.400)
Do you think in that could emerge something that's an approximation of a world model,
Dileep George (1:32:27.600)
which essentially could be used for reasoning? I'm not talking about GPT3, I'm talking about GPT4,
Dileep George (1:32:35.920)
5 and GPT10. Yeah, I mean they will look more impressive than GPT3. So if you take that to
Dileep George (1:32:42.320)
the extreme then a Markov chain of just first order and if you go to, I'm taking the other
Dileep George (1:32:51.520)
extreme, if you read Shannon's book, he has a model of English text which is based on first
Dileep George (1:32:59.200)
order Markov chains, second order Markov chains, third order Markov chains and saying that okay,
Dileep George (1:33:03.120)
third order Markov chains look better than first order Markov chains. So does that mean a first
Dileep George (1:33:09.600)
order Markov chain has a model of the world? Yes, it does. So yes, in that level when you go higher
Dileep George (1:33:18.160)
order models or more sophisticated structure in the model like the transformer networks have,
Dileep George (1:33:24.160)
yes they have a model of the text world, but that is not a model of the world. It's a model
Dileep George (1:33:32.640)
of the text world and it will have interesting properties and it will be useful, but just scaling
Dileep George (1:33:41.120)
it up is not going to give us AGI or natural language understanding or meaning. Well the
Dileep George (1:33:49.280)
question is whether being forced to compress a very large amount of text forces you to construct
Dileep George (1:33:58.880)
things that are very much like, because the ideas of concepts and meaning is a spectrum.
Lex Fridman (1:34:06.800)
Sure, yeah. So in order to form that kind of compression,
Dileep George (1:34:13.920)
maybe it will be forced to figure out abstractions which look awfully a lot like the kind of things
Lex Fridman (1:34:24.160)
that we think about as concepts, as world models, as common sense. Is that possible?
Dileep George (1:34:31.120)
No, I don't think it is possible because the information is not there.
Lex Fridman (1:34:34.320)
The information is there behind the text, right?
Dileep George (1:34:38.640)
No, unless somebody has written down all the details about how everything works in the world
Dileep George (1:34:44.400)
to the absurd amounts like, okay, it is easier to walk forward than backward, that you have to open
Dileep George (1:34:51.040)
the door to go out of the thing, doctors wear underwear. Unless all these things somebody
Dileep George (1:34:56.560)
has written down somewhere or somehow the program found it to be useful for compression from some
Dileep George (1:35:01.680)
other text, the information is not there. So that's an argument that text is a lot
Lex Fridman (1:35:07.840)
lower fidelity than the experience of our physical world.
Dileep George (1:35:13.040)
Right, correct. Pictures worth a thousand words.
Dileep George (1:35:17.440)
Well, in this case, pictures aren't really... So the richest aspect of the physical world isn't
Dileep George (1:35:24.080)
even just pictures, it's the interactivity with the world.
Lex Fridman (1:35:28.240)
Exactly, yeah.
Dileep George (1:35:29.200)
It's being able to interact. It's almost like...
Dileep George (1:35:36.720)
It's almost like if you could interact... Well, maybe I agree with you that pictures
Dileep George (1:35:42.880)
worth a thousand words, but a thousand...
Lex Fridman (1:35:45.760)
It's still... Yeah, you could capture it with the GPTX.
Lex Fridman (1:35:49.760)
So I wonder if there's some interactive element where a system could live in text world where it
Dileep George (1:35:54.400)
could be part of the chat, be part of talking to people. It's interesting. I mean, fundamentally...
Lex Fridman (1:36:03.040)
So you're making a statement about the limitation of text. Okay, so let's say we have a text
Dileep George (1:36:10.960)
corpus that includes basically every experience we could possibly have. I mean, just a very large
Dileep George (1:36:19.280)
corpus of text and also interactive components. I guess the question is whether the neural network
Dileep George (1:36:25.440)
architecture, these very simple transformers, but if they had like hundreds of trillions or
Dileep George (1:36:33.200)
whatever comes after a trillion parameters, whether that could store the information
Dileep George (1:36:42.080)
needed, that's architecturally. Do you have thoughts about the limitation on that side of
Dileep George (1:36:46.880)
things with neural networks? I mean, so transformers are still a feed forward neural
Dileep George (1:36:52.160)
network. It has a very interesting architecture, which is good for text modeling and probably some
Dileep George (1:36:59.200)
aspects of video modeling, but it is still a feed forward architecture. You believe in the
Dileep George (1:37:04.560)
feedback mechanism, the recursion. Oh, and also causality, being able to do counterfactual
Dileep George (1:37:11.280)
reasoning, being able to do interventions, which is actions in the world. So all those things
Dileep George (1:37:20.080)
require different kinds of models to be built. I don't think transformers captures that family. It
Dileep George (1:37:28.400)
is very good at statistical modeling of text and it will become better and better with more data,
Dileep George (1:37:35.280)
bigger models, but that is only going to get so far. So I had this joke on Twitter saying that,
Dileep George (1:37:44.240)
hey, this is a model that has read all of quantum mechanics and theory of relativity and we are
Dileep George (1:37:51.600)
asking you to do text completion or we are asking you to solve simple puzzles. When you have AGI,
Dileep George (1:37:59.280)
that is not what you ask the system to do. We will ask the system to do experiments and come
Dileep George (1:38:08.240)
up with hypothesis and revise the hypothesis based on evidence from experiments, all those things.
Dileep George (1:38:13.680)
Those are the things that we want the system to do when we have AGI, not solve simple puzzles.
Lex Fridman (1:38:20.000)
Like impressive demos, somebody generating a red button in HTML.
Dileep George (1:38:24.080)
Right, which are all useful. There is no dissing the usefulness of it.
Lex Fridman (1:38:29.920)
So by the way, I am playing a little bit of a devil's advocate, so calm down internet.
Lex Fridman (1:38:37.280)
So I am curious almost in which ways will a dumb but large neural network will surprise us.
Dileep George (1:38:47.040)
I completely agree with your intuition. It is just that I do not want to dogmatically
Dileep George (1:38:58.400)
100% put all the chips there. We have been surprised so much. Even the current GPT2 and
Dileep George (1:39:06.160)
GPT3 are so surprising. The self play mechanisms of AlphaZero are really surprising. The fact that
Dileep George (1:39:18.640)
reinforcement learning works at all to me is really surprising. The fact that neural networks work at
Dileep George (1:39:23.440)
all is quite surprising given how nonlinear the space is, the fact that it is able to find local
Dileep George (1:39:30.320)
minima that are at all reasonable. It is very surprising. I wonder sometimes whether us humans
Dileep George (1:39:39.760)
just want for AGI not to be such a dumb thing. Because exactly what you are saying is like
Dileep George (1:39:52.560)
the ideas of concepts and be able to reason with those concepts and connect those concepts in
Dileep George (1:39:57.600)
hierarchical ways and then to be able to have world models. Just everything we are describing
Dileep George (1:40:05.360)
in human language in this poetic way seems to make sense. That is what intelligence and reasoning
Dileep George (1:40:11.120)
are like. I wonder if at the core of it, it could be much dumber. Well, finally it is still
Dileep George (1:40:17.680)
connections and messages passing over. So in that way it is dumb. So I guess the recursion,
Lex Fridman (1:40:24.880)
the feedback mechanism, that does seem to be a fundamental kind of thing.
Dileep George (1:40:32.560)
The idea of concepts. Also memory. Correct. Having an episodic memory. That seems to be
Dileep George (1:40:39.920)
an important thing. So how do we get memory? So we have another piece of work which came
Dileep George (1:40:45.760)
out recently on how do you form episodic memories and form abstractions from them.
Lex Fridman (1:40:52.080)
And we haven't figured out all the connections of that to the overall cognitive architecture.
Lex Fridman (1:40:57.680)
But what are your ideas about how you could have episodic memory? So at least it is very clear
Dileep George (1:41:04.720)
that you need to have two kinds of memory. That is very, very clear. There are things that happen
Dileep George (1:41:13.600)
as statistical patterns in the world, but then there is the one timeline of things that happen
Dileep George (1:41:19.760)
only once in your life. And this day is not going to happen ever again. And that needs to be stored
Dileep George (1:41:27.360)
as just a stream of strings. This is my experience. And then the question is about
Lex Fridman (1:41:36.000)
how do you take that experience and connect it to the statistical part of it? How do you
Dileep George (1:41:40.880)
now say that, okay, I experienced this thing. Now I want to be careful about similar situations.
Lex Fridman (1:41:47.040)
So you need to be able to index that similarity using your other giants that is the model of the
Dileep George (1:41:57.920)
world that you have learned. Although the situation came from the episode, you need to be able to
Dileep George (1:42:02.000)
index the other one. So the episodic memory being implemented as an indexing over the other model
Dileep George (1:42:13.200)
that you're building. So the memories remain and they're indexed into the statistical thing
Dileep George (1:42:24.000)
that you form. Yeah, statistical causal structural model that you built over time. So it's basically
Dileep George (1:42:30.560)
the idea is that the hippocampus is just storing or sequencing a set of pointers that happens over
Dileep George (1:42:41.360)
time. And then whenever you want to reconstitute that memory and evaluate the different aspects of
Dileep George (1:42:48.880)
it, whether it was good, bad, do I need to encounter the situation again? You need the cortex
Dileep George (1:42:55.200)
to reinstantiate, to replay that memory. So how do you find that memory? Like which
Dileep George (1:43:00.880)
direction is the important direction? Both directions are again, bidirectional.
Dileep George (1:43:05.760)
I mean, I guess how do you retrieve the memory? So this is again, hypothesis. We're making this
Dileep George (1:43:11.840)
up. So when you come to a new situation, your cortex is doing inference over in the new situation.
Lex Fridman (1:43:21.200)
And then of course, hippocampus is connected to different parts of the cortex and you have this
Dileep George (1:43:27.600)
deja vu situation, right? Okay, I have seen this thing before. And then in the hippocampus, you can
Dileep George (1:43:35.680)
have an index of, okay, this is when it happened as a timeline. And then you can use the hippocampus
Dileep George (1:43:44.480)
to drive the similar timelines to say now I am, rather than being driven by my current input
Dileep George (1:43:52.240)
stimuli, I am going back in time and rewinding my experience from there, putting back into the
Dileep George (1:43:58.400)
cortex. And then putting it back into the cortex of course affects what you're going to see next
Dileep George (1:44:03.680)
in your current situation. Got it. Yeah. So that's the whole thing, having a world model and then
Dileep George (1:44:09.280)
yeah, connecting to the perception. Yeah, it does seem to be that that's what's happening. On the
Dileep George (1:44:16.320)
neural network side, it's interesting to think of how we actually do that. Yeah. To have a knowledge
Dileep George (1:44:24.240)
base. Yes. It is possible that you can put many of these structures into neural networks and we will
Dileep George (1:44:31.120)
find ways of combining properties of neural networks and graphical models. So, I mean,
Dileep George (1:44:39.440)
it's already started happening. Graph neural networks are kind of a merge between them.
Dileep George (1:44:43.840)
Yeah. And there will be more of that thing. So, but to me it is, the direction is pretty clear,
Dileep George (1:44:51.440)
looking at biology and the history of evolutionary history of intelligence, it is pretty clear that,
Dileep George (1:44:59.600)
okay, what is needed is more structure in the models and modeling of the world and supporting
Dileep George (1:45:06.480)
dynamic inference. Well, let me ask you, there's a guy named Elon Musk, there's a company called
Dileep George (1:45:13.600)
Neuralink and there's a general field called brain computer interfaces. Yeah. It's kind of a
Dileep George (1:45:20.480)
interface between your two loves. Yes. The brain and the intelligence. So there's like
Dileep George (1:45:26.560)
very direct applications of brain computer interfaces for people with different conditions,
Dileep George (1:45:32.160)
more in the short term. Yeah. But there's also these sci fi futuristic kinds of ideas of AI
Dileep George (1:45:38.320)
systems being able to communicate in a high bandwidth way with the brain, bidirectional.
Dileep George (1:45:45.600)
Yeah. What are your thoughts about Neuralink and BCI in general as a possibility? So I think BCI
Dileep George (1:45:53.840)
is a cool research area. And in fact, when I got interested in brains initially, when I was
Dileep George (1:46:02.240)
enrolled at Stanford and when I got interested in brains, it was through a brain computer
Dileep George (1:46:07.840)
interface talk that Krishna Shenoy gave. That's when I even started thinking about the problem.
Lex Fridman (1:46:14.160)
So it is definitely a fascinating research area and the applications are enormous. So there is a
Dileep George (1:46:21.200)
science fiction scenario of brains directly communicating. Let's keep that aside for the
Dileep George (1:46:26.160)
time being. Even just the intermediate milestones that pursuing, which are very reasonable as far
Dileep George (1:46:32.400)
as I can see, being able to control an external limb using direct connections from the brain
Lex Fridman (1:46:40.560)
and being able to write things into the brain. So those are all good steps to take and they have
Dileep George (1:46:49.120)
enormous applications. People losing limbs being able to control prosthetics, quadriplegics being
Dileep George (1:46:55.280)
able to control something, and therapeutics. I also know about another company working in
Dileep George (1:47:01.440)
the space called Paradromics. They're based on a different electrode array, but trying to attack
Dileep George (1:47:09.120)
some of the same problems. So I think it's a very... Also surgery? Correct. Surgically implanted
Dileep George (1:47:14.800)
electrodes. Yeah. So yeah, I think of it as a very, very promising field, especially when it is
Dileep George (1:47:22.560)
helping people overcome some limitations. Now, at some point, of course, it will advance the level of
Dileep George (1:47:29.040)
being able to communicate. How hard is that problem do you think? Let's say we magically solve
Lex Fridman (1:47:37.440)
what I think is a really hard problem of doing all of this safely. Yeah. So being able to connect
Dileep George (1:47:45.600)
electrodes and not just thousands, but like millions to the brain. I think it's very,
Dileep George (1:47:51.440)
very hard because you also do not know what will happen to the brain with that in the sense of how
Dileep George (1:47:58.160)
does the brain adapt to something like that? And as we were learning, the brain is quite,
Dileep George (1:48:04.800)
in terms of neuroplasticity, is pretty malleable. Correct. So it's going to adjust. Correct. So the
Dileep George (1:48:10.480)
machine learning side, the computer side is going to adjust, and then the brain is going to adjust.
Dileep George (1:48:14.480)
Exactly. And then what soup does this land us into? The kind of hallucinations you might get
Dileep George (1:48:20.400)
from this that might be pretty intense. Just connecting to all of Wikipedia. It's interesting
Dileep George (1:48:28.080)
whether we need to be able to figure out the basic protocol of the brain's communication schemes
Dileep George (1:48:34.960)
in order to get them to the machine and the brain to talk. Because another possibility is the brain
Dileep George (1:48:41.120)
actually just adjust to whatever the heck the computer is doing. Exactly. That's the way I think
Dileep George (1:48:45.280)
that I find that to be a more promising way. It's basically saying, okay, attach electrodes
Dileep George (1:48:51.440)
to some part of the cortex. Maybe if it is done from birth, the brain will adapt. It says that
Dileep George (1:48:58.880)
that part is not damaged. It was not used for anything. These electrodes are attached there.
Lex Fridman (1:49:02.880)
And now you train that part of the brain to do this high bandwidth communication between
Dileep George (1:49:09.120)
something else. And if you do it like that, then it is brain adapting to... And of course,
Dileep George (1:49:15.680)
your external system is designed so that it is adaptable. Just like we designed computers
Dileep George (1:49:21.200)
or mouse, keyboard, all of them to be interacting with humans. So of course, that feedback system
Dileep George (1:49:28.720)
is designed to be human compatible, but now it is not trying to record from all of the brain.
Lex Fridman (1:49:37.360)
And now two systems trying to adapt to each other. It's the brain adapting into one way.
Dileep George (1:49:44.160)
That's fascinating. The brain is connected to the internet. Just imagine just connecting it
Dileep George (1:49:51.520)
to Twitter and just taking that stream of information. Yeah. But again, if we take a
Dileep George (1:49:59.760)
step back, I don't know what your intuition is. I feel like that is not as hard of a problem as the
Dileep George (1:50:08.720)
doing it safely. There's a huge barrier to surgery because the biological system, it's a mush of
Dileep George (1:50:19.200)
like weird stuff. So that the surgery part of it, biology part of it, the longterm repercussions
Dileep George (1:50:26.800)
part of it. I don't know what else will... We often find after a long time in biology that,
Dileep George (1:50:35.440)
okay, that idea was wrong. So people used to cut off the gland called the thymus or something.
Lex Fridman (1:50:43.680)
And then they found that, oh no, that actually causes cancer.
Lex Fridman (1:50:50.560)
And then there's a subtle like millions of variables involved. But this whole process,
Dileep George (1:50:55.440)
the nice thing, just like again with Elon, just like colonizing Mars, seems like a ridiculously
Dileep George (1:51:02.000)
difficult idea. But in the process of doing it, we might learn a lot about the biology of the
Dileep George (1:51:08.320)
neurobiology of the brain, the neuroscience side of things. It's like, if you want to learn
Dileep George (1:51:13.520)
something, do the most difficult version of it and see what you learn. The intermediate steps
Dileep George (1:51:19.520)
that they are taking sounded all very reasonable to me. It's great. Well, but like everything with
Dileep George (1:51:25.680)
Elon is the timeline seems insanely fast. So that's the only awful question. Well,
Lex Fridman (1:51:34.000)
we've been talking about cognition a little bit. So like reasoning,
Lex Fridman (1:51:38.640)
we haven't mentioned the other C word, which is consciousness. Do you ever think about that one?
Dileep George (1:51:43.840)
Is that useful at all in this whole context of what it takes to create an intelligent reasoning
Lex Fridman (1:51:51.520)
being? Or is that completely outside of your, like the engineering perspective of intelligence?
Dileep George (1:51:58.400)
It is not outside the realm, but it doesn't on a day to day basis inform what we do,
Lex Fridman (1:52:05.120)
but it's more, so in many ways, the company name is connected to this idea of consciousness.
Dileep George (1:52:12.160)
What's the company name? Vicarious. So Vicarious is the company name. And so what does Vicarious
Dileep George (1:52:19.600)
mean? At the first level, it is about modeling the world and it is internalizing the external actions.
Lex Fridman (1:52:29.360)
So you interact with the world and learn a lot about the world. And now after having learned
Dileep George (1:52:34.960)
a lot about the world, you can run those things in your mind without actually having to act
Dileep George (1:52:42.080)
in the world. So you can run things vicariously just in your brain. And similarly, you can
Dileep George (1:52:48.800)
experience another person's thoughts by having a model of how that person works
Lex Fridman (1:52:54.560)
and running there, putting yourself in some other person's shoes. So that is being vicarious.
Dileep George (1:53:01.280)
Now it's the same modeling apparatus that you're using to model the external world
Dileep George (1:53:06.800)
or some other person's thoughts. You can turn it to yourself. If that same modeling thing is
Dileep George (1:53:14.320)
applied to your own modeling apparatus, then that is what gives rise to consciousness, I think.
Dileep George (1:53:21.040)
Well, that's more like self awareness. There's the hard problem of consciousness, which is
Dileep George (1:53:25.840)
when the model feels like something, when this whole process is like you really are in it.
Dileep George (1:53:37.680)
You feel like an entity in this world. Not just you know that you're an entity, but it feels like
Dileep George (1:53:43.920)
something to be that entity. And thereby, we attribute this. Then it starts to be where
Dileep George (1:53:54.400)
something that has consciousness can suffer. You start to have these kinds of things that we can
Dileep George (1:53:59.120)
reason about that is much heavier. It seems like there's much greater cost to your decisions.
Lex Fridman (1:54:09.520)
And mortality is tied up into that. The fact that these things end. First of all, I end at some
Dileep George (1:54:18.640)
point, and then other things end. That somehow seems to be, at least for us humans, a deep
Dileep George (1:54:27.840)
motivator. That idea of motivation in general, we talk about goals in AI, but goals aren't quite
Dileep George (1:54:38.320)
the same thing as our mortality. It feels like, first of all, humans don't have a goal, and they
Dileep George (1:54:46.560)
just kind of create goals at different levels. They make up goals because we're terrified by
Dileep George (1:54:54.240)
the mystery of the thing that gets us all. We make these goals up. We're like a goal generation
Dileep George (1:55:02.880)
machine, as opposed to a machine which optimizes the trajectory towards a singular goal. It feels
Dileep George (1:55:10.880)
like that's an important part of cognition, that whole mortality thing. Well, it is a part of human
Dileep George (1:55:18.480)
cognition, but there is no reason for that mortality to come to the equation for an artificial
Dileep George (1:55:30.080)
system, because we can copy the artificial system. The problem with humans is that I can't clone
Dileep George (1:55:36.800)
you. Even if I clone you as the hardware, your experience that was stored in your brain,
Dileep George (1:55:45.760)
your episodic memory, all those will not be captured in the new clone. But that's not the
Dileep George (1:55:52.880)
same with an AI system. But it's also possible that the thing that you mentioned with us humans
Dileep George (1:56:02.320)
is actually of fundamental importance for intelligence. The fact that you can copy an AI
Dileep George (1:56:07.760)
system means that that AI system is not yet an AGI. If you look at existence proof, if we reason
Dileep George (1:56:18.240)
based on existence proof, you could say that it doesn't feel like death is a fundamental property
Dileep George (1:56:24.080)
of an intelligent system. But we don't yet. Give me an example of an immortal intelligent being.
Dileep George (1:56:33.840)
We don't have those. It's very possible that that is a fundamental property of intelligence,
Dileep George (1:56:42.240)
is a thing that has a deadline for itself. Well, you can think of it like this. Suppose you invent
Dileep George (1:56:49.840)
a way to freeze people for a long time. It's not dying. So you can be frozen and woken up
Dileep George (1:56:58.160)
thousands of years from now. So it's no fear of death. Well, no, it's not about time. It's about
Dileep George (1:57:08.000)
the knowledge that it's temporary. And that aspect of it, the finiteness of it, I think
Dileep George (1:57:17.120)
creates a kind of urgency. Correct. For us, for humans. Yeah, for humans. Yes. And that is part
Dileep George (1:57:23.200)
of our drives. And that's why I'm not too worried about AI having motivations to kill all humans
Lex Fridman (1:57:35.040)
and those kinds of things. Why? Just wait. So why do you need to do that? I've never heard that
Dileep George (1:57:43.440)
before. That's a good point. Yeah, just murder seems like a lot of work. Let's just wait it out.
Dileep George (1:57:52.560)
They'll probably hurt themselves. Let me ask you, people often kind of wonder, world class researchers
Dileep George (1:58:01.440)
such as yourself, what kind of books, technical fiction, philosophical, had an impact on you and
Dileep George (1:58:10.320)
your life and maybe ones you could possibly recommend that others read? Maybe if you have
Dileep George (1:58:17.920)
three books that pop into mind. Yeah. So I definitely liked Judea Pearl's book,
Dileep George (1:58:23.920)
Probabilistic Reasoning and Intelligent Systems. It's a very deep technical book. But what I liked
Dileep George (1:58:30.640)
is that, so there are many places where you can learn about probabilistic graphical models from.
Lex Fridman (1:58:36.400)
But throughout this book, Judea Pearl kind of sprinkles his philosophical observations and he
Dileep George (1:58:42.960)
thinks about, connects us to how the brain thinks and attentions and resources, all those things. So
Dileep George (1:58:48.400)
that whole thing makes it more interesting to read. He emphasizes the importance of causality.
Lex Fridman (1:58:54.400)
So that was in his later book. So this was the first book, Probabilistic Reasoning and Intelligent
Dileep George (1:58:58.800)
Systems. He mentions causality, but he hadn't really sunk his teeth into causality. But he
Dileep George (1:59:05.040)
really sunk his teeth into, how do you actually formalize it? And the second book,
Lex Fridman (1:59:11.360)
Causality, the one in 2000, that one is really hard. So I would recommend that.
Dileep George (1:59:17.840)
Yeah. So that looks at the mathematical, his model of...
Lex Fridman (1:59:22.560)
Do calculus.
Dileep George (1:59:23.120)
Do calculus. Yeah. It was pretty dense mathematically.
Lex Fridman (1:59:25.520)
Right. The book of Y is definitely more enjoyable.
Dileep George (1:59:28.880)
For sure.
Lex Fridman (1:59:29.360)
Yeah. So I would recommend Probabilistic Reasoning and Intelligent Systems.
Dileep George (1:59:34.160)
Another book I liked was one from Doug Hofstadter. This was a long time ago. He had a book,
Dileep George (1:59:41.360)
I think it was called The Mind's Eye. It was probably Hofstadter and Daniel Dennett together.
Dileep George (1:59:49.200)
Yeah. And I actually was, I bought that book. It's on my show. I haven't read it yet,
Lex Fridman (1:59:54.880)
but I couldn't get an electronic version of it, which is annoying because you read everything on
Lex Fridman (20:02.720)
What happened to V3?
Dileep George (20:03.920)
Well, yeah, that's another pathway. Okay. So this is, this, I'm talking about just object
Dileep George (20:08.880)
recognition pathway.
Lex Fridman (20:09.920)
All right, cool.
Lex Fridman (20:10.880)
And then in V1 itself, so it's, there is a very detailed microcircuit in V1 itself. That is,
Dileep George (20:19.120)
there is organization within a level itself. The cortical sheet is organized into, you know,
Dileep George (20:25.040)
multiple layers and there are columnar structure. And, and this, this layer wise and columnar
Dileep George (20:31.440)
structure is repeated in V1, V2, V4, IT, all of them, right? And, and the connections between
Dileep George (20:38.800)
these layers within a level, you know, in V1 itself, there are six layers roughly, and the
Dileep George (20:44.480)
connections between them, there is a particular structure to them. And now, so one example
Dileep George (20:51.200)
of an experiment people did is when I, when you present a stimulus, which is, let's say,
Dileep George (21:00.400)
requires separating the foreground from the background of an object. So it is, it's a
Dileep George (21:06.240)
textured triangle on a textured background. And you can check, does the surface settle
Lex Fridman (21:14.880)
first or does the contour settle first?
Lex Fridman (21:19.040)
Settle?
Dileep George (21:19.600)
Settle in the sense that the, so when you finally form the percept of the, of the triangle,
Dileep George (21:28.080)
you understand where the contours of the triangle are, and you also know where the inside of
Dileep George (21:32.720)
the triangle is, right? That's when you form the final percept. Now you can ask, what is
Dileep George (21:39.200)
the dynamics of forming that final percept? Do the, do the neurons first find the edges
Lex Fridman (21:48.880)
and converge on where the edges are, and then they find the inner surfaces, or does it go
Lex Fridman (21:55.120)
the other way around?
Lex Fridman (21:55.600)
The other way around. So what's the answer?
Dileep George (21:58.320)
In this case, it turns out that it first settles on the edges. It converges on the edge hypothesis
Lex Fridman (22:05.280)
first, and then the surfaces are filled in from the edges to the inside.
Dileep George (22:10.880)
That's fascinating.
Lex Fridman (22:12.000)
And the detail to which you can study this, it's amazing that you can actually not only
Dileep George (22:18.640)
find the temporal dynamics of when this happens, and then you can also find which layer in
Dileep George (22:25.520)
the, you know, in V1, which layer is encoding the edges, which layer is encoding the surfaces,
Lex Fridman (22:32.960)
and which layer is encoding the feedback, which layer is encoding the feed forward,
Lex Fridman (22:37.440)
and what's the combination of them that produces the final percept.
Lex Fridman (22:42.000)
And these kinds of experiments stand out when you try to explain illusions. One example
Dileep George (22:48.400)
of a favorite illusion of mine is the Kanitsa triangle. I don't know that you are familiar
Dileep George (22:51.920)
with this one. So this is an example where it's a triangle, but only the corners of the
Lex Fridman (23:00.960)
triangle are shown in the stimulus. So they look like kind of Pacman.
Dileep George (23:06.080)
Oh, the black Pacman.
Lex Fridman (23:07.600)
Exactly.
Lex Fridman (23:08.640)
And then you start to see.
Dileep George (23:10.000)
Your visual system hallucinates the edges. And when you look at it, you will see a faint
Dileep George (23:16.400)
edge. And you can go inside the brain and look, do actually neurons signal the presence
Dileep George (23:24.160)
of this edge? And if they signal, how do they do it? Because they are not receiving anything
Dileep George (23:30.320)
from the input. The input is blank for those neurons. So how do they signal it? When does
Dileep George (23:37.840)
the signaling happen? So if a real contour is present in the input, then the neurons
Dileep George (23:45.440)
immediately signal, okay, there is an edge here. When it is an illusory edge, it is clearly
Dileep George (23:52.400)
not in the input. It is coming from the context. So those neurons fire later. And you can say
Dileep George (23:58.720)
that, okay, it's the feedback connection that is causing them to fire. And they happen later.
Lex Fridman (24:05.920)
And I'll find the dynamics of them. So these studies are pretty impressive and very detailed.
Lex Fridman (24:13.280)
So by the way, just a step back, you said that there may be more feedback connections
Dileep George (24:20.080)
than feed forward connections. First of all, if it's just for like a machine learning folks,
Dileep George (24:27.360)
I mean, that's crazy that there's all these feedback connections. We often think about,
Dileep George (24:36.400)
thanks to deep learning, you start to think about the human brain as a kind of feed forward
Dileep George (24:42.720)
mechanism. So what the heck are these feedback connections? What's the dynamics? What are we
Dileep George (24:52.960)
supposed to think about them? So this fits into a very beautiful picture about how the brain works.
Lex Fridman (24:59.360)
So the beautiful picture of how the brain works is that our brain is building a model of the world.
Dileep George (25:06.080)
I know. So our visual system is building a model of how objects behave in the world. And we are
Dileep George (25:13.920)
constantly projecting that model back onto the world. So what we are seeing is not just a feed
Dileep George (25:20.240)
forward thing that just gets interpreted in a feed forward part. We are constantly projecting
Dileep George (25:25.280)
our expectations onto the world. And what the final person is a combination of what we project
Dileep George (25:31.600)
onto the world combined with what the actual sensory input is. Almost like trying to calculate
Dileep George (25:37.920)
the difference and then trying to interpret the difference. Yeah. I wouldn't put this calculating
Dileep George (25:44.000)
the difference. It's more like what is the best explanation for the input stimulus based on the
Dileep George (25:50.640)
model of the world I have. Got it. And that's where all the illusions come in. But that's an
Dileep George (25:56.560)
incredibly efficient process. So the feedback mechanism, it just helps you constantly. Yeah.
Lex Fridman (26:05.360)
So hallucinate how the world should be based on your world model and then just looking at
Dileep George (26:11.680)
if there's novelty, like trying to explain it. Hence, that's why movement. We detect movement
Dileep George (26:19.680)
really well. There's all these kinds of things. And this is like at all different levels of the
Dileep George (26:25.360)
cortex you're saying. This happens at the lowest level or the highest level. Yes. Yeah. In fact,
Dileep George (26:30.480)
feedback connections are more prevalent in everywhere in the cortex. And so one way to
Dileep George (26:36.640)
think about it, and there's a lot of evidence for this, is inference. So basically, if you have a
Dileep George (26:42.800)
model of the world and when some evidence comes in, what you are doing is inference. You are trying
Dileep George (26:50.160)
to now explain this evidence using your model of the world. And this inference includes projecting
Dileep George (26:58.240)
your model onto the evidence and taking the evidence back into the model and doing an
Dileep George (27:04.720)
iterative procedure. And this iterative procedure is what happens using the feed forward feedback
Dileep George (27:11.840)
propagation. And feedback affects what you see in the world, and it also affects feed forward
Dileep George (27:17.680)
propagation. And examples are everywhere. We see these kinds of things everywhere. The idea that
Dileep George (27:25.840)
there can be multiple competing hypotheses in our model trying to explain the same evidence,
Lex Fridman (27:32.480)
and then you have to kind of make them compete. And one hypothesis will explain away the other
Dileep George (27:39.440)
hypothesis through this competition process. So you have competing models of the world
Lex Fridman (27:46.800)
that try to explain. What do you mean by explain away?
Lex Fridman (27:50.000)
So this is a classic example in graphical models, probabilistic models.
Lex Fridman (27:56.800)
What are those?
Dileep George (28:01.120)
I think it's useful to mention because we'll talk about them more.
Lex Fridman (28:05.120)
So neural networks are one class of machine learning models. You have distributed set of
Dileep George (28:12.800)
nodes, which are called the neurons. Each one is doing a dot product and you can approximate
Dileep George (28:18.160)
any function using this multilevel network of neurons. So that's a class of models which are
Dileep George (28:24.720)
useful for function approximation. There is another class of models in machine learning
Dileep George (28:30.480)
called probabilistic graphical models. And you can think of them as each node in that model is
Dileep George (28:38.800)
variable, which is talking about something. It can be a variable representing, is an edge present
Dileep George (28:46.160)
in the input or not? And at the top of the network, a node can be representing, is there an object
Dileep George (28:56.000)
present in the world or not? So it is another way of encoding knowledge. And then once you
Dileep George (29:06.960)
encode the knowledge, you can do inference in the right way. What is the best way to
Dileep George (29:15.280)
explain some set of evidence using this model that you encoded? So when you encode the model,
Dileep George (29:20.880)
you are encoding the relationship between these different variables. How is the edge
Lex Fridman (29:24.800)
connected to the model of the object? How is the surface connected to the model of the object?
Lex Fridman (29:29.600)
And then, of course, this is a very distributed, complicated model. And inference is, how do you
Dileep George (29:37.120)
explain a piece of evidence when a set of stimulus comes in? If somebody tells me there is a 50%
Dileep George (29:42.960)
probability that there is an edge here in this part of the model, how does that affect my belief
Dileep George (29:47.840)
on whether I should think that there is a square present in the image? So this is the process of
Dileep George (29:54.960)
inference. So one example of inference is having this expiring away effect between multiple causes.
Dileep George (2:00:00.800)
Kindle. So you had to actually purchase the physical. It's one of the only physical books
Lex Fridman (2:00:06.560)
I have because anyway, a lot of people recommended it highly. So yeah.
Lex Fridman (2:00:11.200)
And the third one I would definitely recommend reading is, this is not a technical book. It is
Dileep George (2:00:18.720)
history. The name of the book, I think, is Bishop's Boys. It's about Wright brothers
Lex Fridman (2:00:25.040)
and their path and how it was... There are multiple books on this topic and all of them
Dileep George (2:00:34.560)
are great. It's fascinating how flight was treated as an unsolvable problem. And also,
Lex Fridman (2:00:46.400)
what aspects did people emphasize? People thought, oh, it is all about
Dileep George (2:00:51.520)
just powerful engines. You just need to have powerful lightweight engines. And so some people
Dileep George (2:01:00.160)
thought of it as, how far can we just throw the thing? Just throw it.
Lex Fridman (2:01:04.000)
Like a catapult.
Dileep George (2:01:05.040)
Yeah. So it's very fascinating. And even after they made the invention,
Lex Fridman (2:01:11.520)
people are not believing it.
Dileep George (2:01:13.040)
Ah, the social aspect of it.
Lex Fridman (2:01:15.360)
The social aspect. It's very fascinating.
Dileep George (2:01:18.240)
I mean, do you draw any parallels between birds fly? So there's the natural approach to flight
Lex Fridman (2:01:28.320)
and then there's the engineered approach. Do you see the same kind of thing with the brain
Lex Fridman (2:01:33.920)
and our trying to engineer intelligence?
Lex Fridman (2:01:37.280)
Yeah. It's a good analogy to have. Of course, all analogies have their limits.
Lex Fridman (2:01:43.920)
So people in AI often use airplanes as an example of, hey, we didn't learn anything from birds.
Lex Fridman (2:01:55.120)
But the funny thing is that, and the saying is, airplanes don't flap wings. This is what they
Dileep George (2:02:02.560)
say. The funny thing and the ironic thing is that you don't need to flap to fly is something
Dileep George (2:02:09.520)
Wright brothers found by observing birds. So they have in their notebook, in some of these books,
Dileep George (2:02:18.640)
they show their notebook drawings. They make detailed notes about buzzards just soaring over
Dileep George (2:02:26.240)
thermals. And they basically say, look, flapping is not the important, propulsion is not the
Dileep George (2:02:31.440)
important problem to solve here. We want to solve control. And once you solve control,
Dileep George (2:02:37.120)
propulsion will fall into place. All of these are people, they realize this by observing birds.
Dileep George (2:02:44.400)
Beautifully put. That's actually brilliant because people do use that analogy a lot. I'm
Dileep George (2:02:49.280)
going to have to remember that one. Do you have advice for people interested in artificial
Dileep George (2:02:54.480)
intelligence like young folks today? I talk to undergraduate students all the time,
Dileep George (2:02:59.200)
interested in neuroscience, interested in understanding how the brain works. Is there
Lex Fridman (2:03:03.840)
advice you would give them about their career, maybe about their life in general?
Dileep George (2:03:09.520)
Sure. I think every piece of advice should be taken with a pinch of salt, of course,
Dileep George (2:03:14.720)
because each person is different, their motivations are different. But I can definitely
Dileep George (2:03:20.400)
say if your goal is to understand the brain from the angle of wanting to build one, then
Dileep George (2:03:28.480)
being an experimental neuroscientist might not be the way to go about it. A better way to pursue it
Dileep George (2:03:36.240)
might be through computer science, electrical engineering, machine learning, and AI. And of
Dileep George (2:03:42.560)
course, you have to study the neuroscience, but that you can do on your own. If you're more
Dileep George (2:03:48.800)
attracted by finding something intriguing about, discovering something intriguing about the brain,
Dileep George (2:03:53.680)
then of course, it is better to be an experimentalist. So find that motivation,
Lex Fridman (2:03:58.480)
what are you intrigued by? And of course, find your strengths too. Some people are very good
Dileep George (2:04:03.120)
experimentalists and they enjoy doing that. And it's interesting to see which department,
Dileep George (2:04:10.160)
if you're picking in terms of your education path, whether to go with like, at MIT, it's
Dileep George (2:04:18.880)
brain and computer, no, it'd be CS. Yeah. Brain and cognitive sciences, yeah. Or the CS side of
Dileep George (2:04:29.120)
things. And actually the brain folks, the neuroscience folks are more and more now
Dileep George (2:04:34.240)
embracing of learning TensorFlow and PyTorch, right? They see the power of trying to engineer
Dileep George (2:04:44.400)
ideas that they get from the brain into, and then explore how those could be used to create
Dileep George (2:04:52.720)
intelligent systems. So that might be the right department actually. Yeah. So this was a question
Dileep George (2:04:58.640)
in one of the Redwood Neuroscience Institute workshops that Jeff Hawkins organized almost 10
Dileep George (2:05:06.160)
years ago. This question was put to a panel, right? What should be the undergrad major you should
Dileep George (2:05:11.040)
take if you want to understand the brain? And the majority opinion in that one was electrical
Dileep George (2:05:17.200)
engineering. Interesting. Because, I mean, I'm a double undergrad, so I got lucky in that way.
Lex Fridman (2:05:25.040)
But I think it does have some of the right ingredients because you learn about circuits.
Dileep George (2:05:30.080)
You learn about how you can construct circuits to approach, do functions. You learn about
Dileep George (2:05:37.920)
microprocessors. You learn information theory. You learn signal processing. You learn continuous
Dileep George (2:05:43.040)
math. So in that way, it's a good step. If you want to go to computer science or neuroscience,
Lex Fridman (2:05:50.880)
it's a good step. The downside, you're more likely to be forced to use MATLAB.
Dileep George (2:05:56.640)
You're more likely to be forced to use MATLAB. So one of the interesting things about, I mean,
Dileep George (2:06:07.920)
this is changing. The world is changing. But certain departments lagged on the programming
Dileep George (2:06:13.840)
side of things, on developing good habits in terms of software engineering. But I think that's more
Lex Fridman (2:06:19.280)
and more changing. And students can take that into their own hands, like learn to program. I feel
Dileep George (2:06:26.000)
like everybody should learn to program because it, like everyone in the sciences, because it
Dileep George (2:06:34.800)
empowers, it puts the data at your fingertips. So you can organize it. You can find all kinds of
Dileep George (2:06:40.400)
things in the data. And then you can also, for the appropriate sciences, build systems that,
Lex Fridman (2:06:46.240)
like based on that. So like then engineer intelligent systems.
Dileep George (2:06:49.760)
We already talked about mortality. So we hit a ridiculous point. But let me ask you,
Dileep George (2:07:04.800)
one of the things about intelligence is it's goal driven. And you study the brain. So the question
Dileep George (2:07:13.200)
is like, what's the goal that the brain is operating under? What's the meaning of it all
Dileep George (2:07:17.360)
for us humans in your view? What's the meaning of life? The meaning of life is whatever you
Dileep George (2:07:23.920)
construct out of it. It's completely open. It's open. So there's nothing, like you mentioned,
Dileep George (2:07:31.760)
you like constraints. So it's wide open. Is there some useful aspect that you think about in terms
Dileep George (2:07:42.000)
of like the openness of it and just the basic mechanisms of generating goals in studying
Dileep George (2:07:50.480)
cognition in the brain that you think about? Or is it just about, because everything we've talked
Dileep George (2:07:56.640)
about kind of the perception system is to understand the environment. That's like to be
Dileep George (2:08:00.640)
able to like not die, like not fall over and like be able to, you don't think we need to
Dileep George (2:08:09.360)
think about anything bigger than that. Yeah, I think so, because it's basically being able to
Dileep George (2:08:16.160)
understand the machinery of the world such that you can pursue whatever goals you want.
Lex Fridman (2:08:21.600)
So the machinery of the world is really ultimately what we should be striving to understand. The
Lex Fridman (2:08:26.800)
rest is just whatever the heck you want to do or whatever fun you have.
Dileep George (2:08:31.840)
One who is culturally popular. I think that's beautifully put. I don't think there's a better
Dileep George (2:08:42.640)
way to end it. Dilip, I'm so honored that you show up here and waste your time with me. It's
Dileep George (2:08:49.840)
been an awesome conversation. Thanks so much for talking today. Oh, thank you so much. This was
Lex Fridman (2:08:54.400)
so much more fun than I expected. Thank you. Thanks for listening to this conversation with
Dileep George (2:09:00.880)
Dilip George. And thank you to our sponsors, Babbel, Raycon Earbuds, and Masterclass. Please
Dileep George (2:09:07.920)
consider supporting this podcast by going to babbel.com and use code LEX, going to buyraycon.com
Lex Fridman (2:09:16.080)
and signing up at masterclass.com. Click the links, get the discount. It really is the best
Dileep George (2:09:22.240)
way to support this podcast. If you enjoy this thing, subscribe on YouTube, review the Five
Dileep George (2:09:27.440)
Stars Napa podcast, support it on Patreon, or connect with me on Twitter at Lex Friedman,
Dileep George (2:09:33.920)
spelled yes, without the E, just F R I D M A M. And now let me leave you with some words from Marcus
Dileep George (2:09:43.120)
Aurelius. You have power over your mind, not outside events. Realize this and you will find
Lex Fridman (2:09:51.360)
strength. Thank you for listening and hope to see you next time.
Lex Fridman (30:02.080)
So graphical models can be used to represent causality in the world. So let's say, you know,
Dileep George (30:10.800)
your alarm at home can be triggered by a burglar getting into your house, or it can be triggered
Dileep George (30:22.480)
by an earthquake. Both can be causes of the alarm going off. So now, you're in your office,
Dileep George (30:30.640)
you heard burglar alarm going off, you are heading home, thinking that there's a burglar got in. But
Dileep George (30:36.880)
while driving home, if you hear on the radio that there was an earthquake in the vicinity,
Dileep George (30:41.520)
now your strength of evidence for a burglar getting into their house is diminished. Because
Dileep George (30:49.760)
now that piece of evidence is explained by the earthquake being present. So if you think about
Dileep George (30:56.000)
these two causes explaining at lower level variable, which is alarm, now, what we're seeing
Dileep George (31:01.760)
is that increasing the evidence for some cause, you know, there is evidence coming in from below
Dileep George (31:08.000)
for alarm being present. And initially, it was flowing to a burglar being present. But now,
Dileep George (31:14.160)
since there is side evidence for this other cause, it explains away this evidence and evidence will
Dileep George (31:20.800)
now flow to the other cause. This is, you know, two competing causal things trying to explain
Dileep George (31:26.320)
the same evidence. And the brain has a similar kind of mechanism for doing so. That's kind of
Lex Fridman (31:31.840)
interesting. And how's that all encoded in the brain? Like, where's the storage of information?
Dileep George (31:39.280)
Are we talking just maybe to get it a little bit more specific? Is it in the hardware of the actual
Lex Fridman (31:46.160)
connections? Is it in chemical communication? Is it electrical communication? Do we know?
Lex Fridman (31:53.120)
So this is, you know, a paper that we are bringing out soon.
Lex Fridman (31:56.640)
Which one is this?
Dileep George (31:57.680)
This is the cortical microcircuits paper that I sent you a draft of. Of course, this is a lot of
Dileep George (32:03.920)
this. A lot of it is still hypothesis. One hypothesis is that you can think of a cortical column
Dileep George (32:09.840)
as encoding a concept. A concept, you know, think of it as an example of a concept. Is an edge
Dileep George (32:20.800)
present or not? Or is an object present or not? Okay, so you can think of it as a binary variable,
Dileep George (32:27.280)
a binary random variable. The presence of an edge or not, or the presence of an object or not.
Lex Fridman (32:32.000)
So each cortical column can be thought of as representing that one concept, one variable.
Lex Fridman (32:38.080)
And then the connections between these cortical columns are basically encoding the relationship
Dileep George (32:43.680)
between these random variables. And then there are connections within the cortical column.
Dileep George (32:49.360)
Each cortical column is implemented using multiple layers of neurons with very, very,
Dileep George (32:54.320)
very rich structure there. You know, there are thousands of neurons in a cortical column.
Lex Fridman (33:00.240)
But that structure is similar across the different cortical columns.
Dileep George (33:03.520)
Correct. And also these cortical columns connect to a substructure called thalamus.
Lex Fridman (33:10.160)
So all cortical columns pass through this substructure. So our hypothesis is that
Dileep George (33:17.120)
the connections between the cortical columns implement this, you know, that's where the
Dileep George (33:21.600)
knowledge is stored about how these different concepts connect to each other. And then the
Dileep George (33:28.800)
neurons inside this cortical column and in thalamus in combination implement this actual
Dileep George (33:35.760)
computation for inference, which includes explaining away and competing between the
Dileep George (33:41.040)
different hypotheses. And it is all very... So what is amazing is that neuroscientists have
Dileep George (33:49.280)
actually done experiments to the tune of showing these things. They might not be putting it in the
Dileep George (33:55.920)
overall inference framework, but they will show things like, if I poke this higher level neuron,
Dileep George (34:03.120)
it will inhibit through this complicated loop through thalamus, it will inhibit this other
Dileep George (34:07.920)
column. So they will do such experiments. But do they use terminology of concepts,
Dileep George (34:14.080)
for example? So, I mean, is it something where it's easy to anthropomorphize
Lex Fridman (34:22.960)
and think about concepts like you started moving into logic based kind of reasoning systems. So
Dileep George (34:31.200)
I would just think of concepts in that kind of way, or is it a lot messier, a lot more gray area,
Dileep George (34:40.400)
you know, even more gray, even more messy than the artificial neural network kinds,
Dileep George (34:47.200)
kinds of abstractions? Easiest way to think of it as a variable,
Dileep George (34:50.480)
right? It's a binary variable, which is showing the presence or absence of something.
Dileep George (34:55.360)
So, but I guess what I'm asking is, is that something that we're supposed to think of
Lex Fridman (35:01.440)
something that's human interpretable of that something?
Dileep George (35:04.080)
It doesn't need to be. It doesn't need to be human interpretable. There's no need for it to
Dileep George (35:07.920)
be human interpretable. But it's almost like you will be able to find some interpretation of it
Dileep George (35:17.440)
because it is connected to the other things that you know about.
Lex Fridman (35:20.800)
Yeah. And the point is it's useful somehow.
Dileep George (35:23.840)
Yeah. It's useful as an entity in the graphic,
Lex Fridman (35:29.520)
in connecting to the other entities that are, let's call them concepts.
Lex Fridman (35:33.280)
Right. Okay. So, by the way, are these the cortical microcircuits?
Dileep George (35:38.880)
Correct. These are the cortical microcircuits. You know, that's what neuroscientists use to
Dileep George (35:43.120)
talk about the circuits within a level of the cortex. So, you can think of, you know,
Dileep George (35:49.840)
let's think of a neural network, artificial neural network terms. People talk about the
Dileep George (35:54.960)
architecture of how many layers they build, what is the fan in, fan out, et cetera. That is the
Dileep George (36:01.600)
macro architecture. And then within a layer of the neural network, the cortical neural network
Dileep George (36:11.120)
is much more structured within a level. There's a lot more intricate structure there. But even
Dileep George (36:18.160)
within an artificial neural network, you can think of feature detection plus pooling as one
Dileep George (36:23.520)
level. And so, that is kind of a microcircuit. It's much more complex in the real brain. And so,
Dileep George (36:32.880)
within a level, whatever is that circuitry within a column of the cortex and between the layers of
Dileep George (36:38.080)
the cortex, that's the microcircuitry. I love that terminology. Machine learning
Lex Fridman (36:43.040)
people don't use the circuit terminology. Right.
Lex Fridman (36:45.760)
But they should. It's nice. So, okay. Okay. So, that's the cortical microcircuit. So,
Dileep George (36:53.920)
what's interesting about, what can we say, what is the paper that you're working on
Lex Fridman (37:00.640)
propose about the ideas around these cortical microcircuits?
Dileep George (37:04.320)
So, this is a fully functional model for the microcircuits of the visual cortex.
Dileep George (37:10.640)
So, the paper focuses on your idea and our discussion now is focusing on vision.
Lex Fridman (37:15.520)
Yeah. The visual cortex. Okay. So,
Dileep George (37:18.800)
this is a model. This is a full model. This is how vision works.
Lex Fridman (37:22.880)
But this is a hypothesis. Okay. So, let me step back a bit. So, we looked at neuroscience for
Dileep George (37:32.000)
insights on how to build a vision model. Right.
Lex Fridman (37:35.280)
And we synthesized all those insights into a computational model. This is called the recursive
Dileep George (37:40.560)
cortical network model that we used for breaking captures. And we are using the same model for
Lex Fridman (37:47.760)
robotic picking and tracking of objects. And that, again, is a vision system.
Dileep George (37:52.320)
That's a vision system. Computer vision system.
Lex Fridman (37:54.400)
That's a computer vision system. Takes in images and outputs what?
Dileep George (37:59.120)
On one side, it outputs the class of the image and also segments the image. And you can also ask it
Dileep George (38:06.560)
further queries. Where is the edge of the object? Where is the interior of the object? So, it's a
Dileep George (38:11.600)
model that you build to answer multiple questions. So, you're not trying to build a model for just
Dileep George (38:17.120)
classification or just segmentation, et cetera. It's a joint model that can do multiple things.
Dileep George (38:23.440)
So, that's the model that we built using insights from neuroscience. And some of those insights are
Lex Fridman (38:30.080)
what is the role of feedback connections? What is the role of lateral connections? So,
Dileep George (38:34.160)
all those things went into the model. The model actually uses feedback connections.
Lex Fridman (38:38.800)
All these ideas from neuroscience. Yeah.
Dileep George (38:41.440)
So, what the heck is a recursive cortical network? What are the architecture approaches,
Lex Fridman (38:47.200)
interesting aspects here, which is essentially a brain inspired approach to computer vision?
Dileep George (38:54.400)
Yeah. So, there are multiple layers to this question. I can go from the very,
Dileep George (38:58.880)
very top and then zoom in. Okay. So, one important thing, constraint that went into the model is that
Dileep George (39:05.840)
you should not think vision, think of vision as something in isolation. We should not think
Dileep George (39:11.600)
perception as something as a preprocessor for cognition. Perception and cognition are interconnected.
Lex Fridman (39:19.200)
And so, you should not think of one problem in separation from the other problem. And so,
Dileep George (39:24.800)
that means if you finally want to have a system that understand concepts about the world and can
Dileep George (39:30.720)
learn a very conceptual model of the world and can reason and connect to language, all of those
Dileep George (39:36.000)
things, you need to think all the way through and make sure that your perception system
Dileep George (39:41.920)
is compatible with your cognition system and language system and all of them.
Lex Fridman (39:45.920)
And one aspect of that is top down controllability. What does that mean?
Dileep George (39:52.320)
So, that means, you know, so think of, you know, you can close your eyes and think about
Dileep George (39:58.480)
the details of one object, right? I can zoom in further and further. So, think of the bottle in
Dileep George (40:05.600)
front of me, right? And now, you can think about, okay, what the cap of that bottle looks.
Dileep George (40:11.280)
I know we can think about what's the texture on that bottle of the cap. You know, you can think
Dileep George (40:18.000)
about, you know, what will happen if something hits that. So, you can manipulate your visual
Dileep George (40:25.760)
knowledge in cognition driven ways. Yes. And so, this top down controllability and being able to
Dileep George (40:35.520)
simulate scenarios in the world. So, you're not just a passive player in this perception game.
Dileep George (40:43.920)
You can control it. You have imagination. Correct. Correct. So, basically, you know,
Dileep George (40:50.320)
basically having a generative network, which is a model and it is not just some arbitrary
Dileep George (40:56.000)
generative network. It has to be built in a way that it is controllable top down. It is not just
Dileep George (41:02.000)
trying to generate a whole picture at once. You know, it's not trying to generate photorealistic
Dileep George (41:07.760)
things of the world. You know, you don't have good photorealistic models of the world. Human
Dileep George (41:11.520)
brains do not have. If I, for example, ask you the question, what is the color of the letter E
Dileep George (41:17.360)
in the Google logo? You have no idea. Although, you have seen it millions of times, hundreds of
Dileep George (41:25.360)
times. So, it's not, our model is not photorealistic, but it has other properties that we can
Dileep George (41:32.240)
manipulate it. And you can think about filling in a different color in that logo. You can think
Dileep George (41:37.840)
about expanding the letter E. You know, you can see what, so you can imagine the consequence of,
Dileep George (41:44.400)
you know, actions that you have never performed. So, these are the kind of characteristics the
Dileep George (41:49.040)
generative model need to have. So, this is one constraint that went into our model. Like, you
Dileep George (41:52.800)
know, so this is, when you read the, just the perception side of the paper, it is not obvious
Dileep George (41:57.920)
that this was a constraint into the, that went into the model, this top down controllability
Dileep George (42:02.720)
of the generative model. So, what does top down controllability in a model look like? It's a
Dileep George (42:10.480)
really interesting concept. Fascinating concept. What does that, is that the recursiveness gives
Dileep George (42:16.000)
you that? Or how do you do it? Quite a few things. It's like, what does the model factor,
Dileep George (42:22.080)
factorize? You know, what are the, what is the model representing as different pieces in the
Dileep George (42:26.720)
puzzle? Like, you know, so, so in the RCN network, it thinks of the world, you know, so what I said,
Dileep George (42:33.440)
the background of an image is modeled separately from the foreground of the image. So,
Dileep George (42:39.040)
the objects are separate from the background. They are different entities. So, there's a kind
Dileep George (42:43.200)
of segmentation that's built in fundamentally. And then even that object is composed of parts.
Lex Fridman (42:49.840)
And also, another one is the shape of the object is differently modeled from the texture of the
Dileep George (42:57.440)
object. Got it. So, there's like these, you know who Francois Chollet is? Yeah. So, there's, he
Dileep George (43:08.800)
developed this like IQ test type of thing for ARC challenge for, and it's kind of cool that there's
Dileep George (43:16.160)
these concepts, priors that he defines that you bring to the table in order to be able to reason
Dileep George (43:22.560)
about basic shapes and things in IQ test. So, here you're making it quite explicit that here are the
Dileep George (43:30.080)
things that you should be, these are like distinct things that you should be able to model in this.
Dileep George (43:36.960)
Keep in mind that you can derive this from much more general principles. It doesn't, you don't
Dileep George (43:42.240)
need to explicitly put it as, oh, objects versus foreground versus background, the surface versus
Dileep George (43:48.880)
the structure. No, these are, these are derivable from more fundamental principles of how, you know,
Dileep George (43:55.440)
what's the property of continuity of natural signals. What's the property of continuity of
Dileep George (44:01.520)
natural signals? Yeah. By the way, that sounds very poetic, but yeah. So, you're saying that's a,
Dileep George (44:07.920)
there's some low level properties from which emerges the idea that shapes should be different
Dileep George (44:12.560)
than like there should be a parts of an object. There should be, I mean, kind of like Francois,
Dileep George (44:18.640)
I mean, there's objectness, there's all these things that it's kind of crazy that we humans,
Dileep George (44:25.040)
I guess, evolved to have because it's useful for us to perceive the world. Yeah. Correct. And it
Dileep George (44:30.240)
derives mostly from the properties of natural signals. And so, natural signals. So, natural
Dileep George (44:38.080)
signals are the kind of things we'll perceive in the natural world. Correct. I don't know. I don't
Dileep George (44:43.200)
know why that sounds so beautiful. Natural signals. Yeah. As opposed to a QR code, right? Which is an
Dileep George (44:48.080)
artificial signal that we created. Humans are not very good at classifying QR codes. We are very
Dileep George (44:52.880)
good at saying something is a cat or a dog, but not very good at, you know, where computers are
Dileep George (44:58.480)
very good at classifying QR codes. So, our visual system is tuned for natural signals. So,
Dileep George (45:05.600)
it's tuned for natural signals. And there are fundamental assumptions in the architecture
Dileep George (45:11.680)
that are derived from natural signals properties. I wonder when you take hallucinogenic drugs,
Lex Fridman (45:18.640)
does that go into natural or is that closer to the QR code? It's still natural. It's still natural?
Dileep George (45:25.120)
Yeah. Because it is still operating using your brains. By the way, on that topic, I mean,
Dileep George (45:30.480)
I haven't been following. I think they're becoming legalized and certain. I can't wait
Dileep George (45:34.640)
they become legalized to a degree that you, like, vision science researchers could study it.
Dileep George (45:40.080)
Yeah. Just like through medical, chemical ways, modify. There could be ethical concerns, but
Dileep George (45:47.600)
modify. That's another way to study the brain is to be able to chemically modify it. There's
Dileep George (45:53.280)
probably very long a way to figure out how to do it ethically. Yeah, but I think there are studies
Dileep George (46:01.200)
on that already. Yeah, I think so. Because it's not unethical to give it to rats.
Dileep George (46:08.080)
Oh, that's true. That's true. There's a lot of drugged up rats out there. Okay, cool. Sorry.
Dileep George (46:15.600)
Sorry. It's okay. So, there's these low level things from natural signals that...
Dileep George (46:23.840)
...from which these properties will emerge. But it is still a very hard problem on how to encode
Dileep George (46:33.840)
that. So, you mentioned the priors Francho wanted to encode in the abstract reasoning challenge,
Lex Fridman (46:44.880)
but it is not straightforward how to encode those priors. So, some of those challenges,
Dileep George (46:50.960)
like the object completion challenges are things that we purely use our visual system to do.
Dileep George (46:57.840)
It looks like abstract reasoning, but it is purely an output of the vision system. For example,
Dileep George (47:03.200)
completing the corners of that condenser triangle, completing the lines of that condenser triangle.
Dileep George (47:07.120)
It's purely a visual system property. There is no abstract reasoning involved. It uses all these
Dileep George (47:12.160)
priors, but it is stored in our visual system in a particular way that is amenable to inference.
Dileep George (47:18.720)
That is one of the things that we tackled in the... Basically saying, okay, these are the
Dileep George (47:25.440)
prior knowledge which will be derived from the world, but then how is that prior knowledge
Dileep George (47:31.440)
represented in the model such that inference when some piece of evidence comes in can be
Dileep George (47:38.080)
done very efficiently and in a very distributed way? Because there are so many ways of representing
Dileep George (47:44.640)
knowledge, which is not amenable to very quick inference, quick lookups. So that's one core part
Lex Fridman (47:53.840)
of what we tackled in the RCN model. How do you encode visual knowledge to do very quick inference?
Lex Fridman (48:02.800)
Can you maybe comment on... So folks listening to this in general may be familiar with
Lex Fridman (48:08.560)
different kinds of architectures of a neural networks.
Lex Fridman (48:10.720)
What are we talking about with RCN? What does the architecture look like? What are the different
Dileep George (48:16.240)
components? Is it close to neural networks? Is it far away from neural networks? What does it look
Dileep George (48:20.720)
like? Yeah. So you can think of the Delta between the model and a convolutional neural network,
Dileep George (48:27.040)
if people are familiar with convolutional neural networks. So convolutional neural networks have
Dileep George (48:31.440)
this feed forward processing cascade, which is called feature detectors and pooling. And that
Dileep George (48:37.440)
is repeated in a multi level system. And if you want an intuitive idea of what is happening,
Dileep George (48:46.320)
feature detectors are detecting interesting co occurrences in the input. It can be a line,
Dileep George (48:53.920)
a corner, an eye or a piece of texture, et cetera. And the pooling neurons are doing some local
Dileep George (49:03.200)
transformation of that and making it invariant to local transformations. So this is what the
Dileep George (49:07.840)
structure of convolutional neural network is. Recursive cortical network has a similar structure
Dileep George (49:14.880)
when you look at just the feed forward pathway. But in addition to that, it is also structured
Dileep George (49:19.600)
in a way that it is generative so that it can run it backward and combine the forward with the
Dileep George (49:25.680)
backward. Another aspect that it has is it has lateral connections. So if you have an edge here
Lex Fridman (49:37.280)
and an edge here, it has connections between these edges. It is not just feed forward connections.
Dileep George (49:42.080)
It is something between these edges, which is the nodes representing these edges, which is to
Dileep George (49:49.280)
enforce compatibility between them. So otherwise what will happen is that constraints. It's a
Dileep George (49:53.920)
constraint. It's basically if you do just feature detection followed by pooling, then your
Dileep George (50:01.200)
transformations in different parts of the visual field are not coordinated. And so you will create
Dileep George (50:07.760)
a jagged, when you generate from the model, you will create jagged things and uncoordinated
Dileep George (50:14.480)
transformations. So these lateral connections are enforcing the transformations.
Lex Fridman (50:20.160)
Is the whole thing still differentiable?
Lex Fridman (50:22.160)
No, it's not. It's not trained using backprop.
Dileep George (50:27.440)
Okay. That's really important. So there's this feed forward, there's feedback mechanisms.
Dileep George (50:33.280)
There's some interesting connectivity things. It's still layered like multiple layers.
Dileep George (50:41.040)
Okay. Very, very interesting. And yeah. Okay. So the interconnection between adjacent connections
Lex Fridman (50:48.240)
across service constraints that keep the thing stable.
Dileep George (50:52.880)
Correct.
Lex Fridman (50:53.680)
Okay. So what else?
Lex Fridman (50:55.840)
And then there's this idea of doing inference. A neural network does not do inference on the fly.
Lex Fridman (51:03.120)
So an example of why this inference is important is, you know, so one of the first applications
Dileep George (51:09.200)
that we showed in the paper was to crack text based captures.
Lex Fridman (51:15.040)
What are captures?
Dileep George (51:16.000)
I mean, by the way, one of the most awesome, like the people don't use this term anymore
Lex Fridman (51:21.040)
as human computation, I think. I love this term. The guy who created captures,
Lex Fridman (51:26.640)
I think came up with this term. I love it. Anyway. What are captures?
Lex Fridman (51:32.640)
So captures are those things that you fill in when you're, you know, if you're
Dileep George (51:38.480)
opening a new account in Google, they show you a picture, you know, usually
Dileep George (51:43.200)
it used to be set of garbled letters that you have to kind of figure out what is that string
Dileep George (51:48.720)
of characters and type it. And the reason captures exist is because, you know, Google or Twitter
Dileep George (51:56.640)
do not want automatic creation of accounts. You can use a computer to create millions of accounts
Lex Fridman (52:03.200)
and use that for nefarious purposes. So you want to make sure that to the extent possible,
Dileep George (52:10.560)
the interaction that their system is having is with a human. So it's a, it's called a human
Dileep George (52:16.080)
interaction proof. A capture is a human interaction proof. So, so this is a captures are by design,
Lex Fridman (52:23.840)
things that are easy for humans to solve, but hard for computers.
Dileep George (52:27.360)
Hard for robots.
Lex Fridman (52:28.240)
Yeah. So, and text based captures was the one which is prevalent around 2014,
Dileep George (52:36.320)
because at that time, text based captures were hard for computers to crack. Even now,
Dileep George (52:42.240)
they are actually in the sense of an arbitrary text based capture will be unsolvable even now,
Lex Fridman (52:48.240)
but with the techniques that we have developed, it can be, you know, you can quickly develop
Lex Fridman (52:52.320)
a mechanism that solves the capture.
Dileep George (52:55.360)
They've probably gotten a lot harder too. They've been getting clever and clever
Dileep George (53:00.320)
generating these text captures. So, okay. So that was one of the things you've tested it on is these
Lex Fridman (53:06.640)
kinds of captures in 2014, 15, that kind of stuff. So what, I mean, why, by the way, why captures?
Dileep George (53:15.120)
Yeah. Even now, I would say capture is a very, very good challenge problem. If you want to
Dileep George (53:21.920)
understand how human perception works, and if you want to build systems that work,
Dileep George (53:27.040)
like the human brain, and I wouldn't say capture is a solved problem. We have cracked the fundamental
Dileep George (53:32.880)
defense of captures, but it is not solved in the way that humans solve it. So I can give an example.
Dileep George (53:40.000)
I can take a five year old child who has just learned characters and show them any new capture
Dileep George (53:48.640)
that we create. They will be able to solve it. I can show you, I can show you a picture of a
Dileep George (53:56.400)
character. I can show you pretty much any new capture from any new website. You'll be able to
Dileep George (54:02.000)
solve it without getting any training examples from that particular style of capture.
Lex Fridman (54:06.640)
You're assuming I'm human. Yeah.
Dileep George (54:08.000)
Yes. Yeah. That's right. So if you are human, otherwise I will be able to figure that out
Dileep George (54:15.440)
using this one. But this whole podcast is just a touring test, a long touring test. Anyway,
Dileep George (54:22.000)
yeah. So humans can figure it out with very few examples. Or no training examples. No training
Dileep George (54:28.880)
examples from that particular style of capture. So even now this is unreachable for the current
Dileep George (54:37.760)
deep learning system. So basically there is no, I don't think a system exists where you can
Dileep George (54:41.760)
basically say, train on whatever you want. And then now say, hey, I will show you a new capture,
Dileep George (54:47.840)
which I did not show you in the training setup. Will the system be able to solve it? It still
Dileep George (54:54.160)
doesn't exist. So that is the magic of human perception. And Doug Hofstadter put this very
Dileep George (55:01.760)
beautifully in one of his talks. The central problem in AI is what is the letter A. If you
Dileep George (55:11.440)
can build a system that reliably can detect all the variations of the letter A, you don't even
Dileep George (55:17.600)
know to go to the B and the C. Yeah. You don't even know to go to the B and the C or the strings
Lex Fridman (55:23.040)
of characters. And so that is the spirit with which we tackle that problem.
Lex Fridman (55:28.880)
What does it mean by that? I mean, is it like without training examples, try to figure out
Lex Fridman (55:36.160)
the fundamental elements that make up the letter A in all of its forms?
Dileep George (55:43.520)
In all of its forms. A can be made with two humans standing, leaning against each other,
Lex Fridman (55:47.920)
holding the hands. And it can be made of leaves.
Dileep George (55:52.080)
Yeah. You might have to understand everything about this world in order to understand the
Lex Fridman (55:56.480)
letter A. Yeah. Exactly.
Lex Fridman (55:57.920)
So it's common sense reasoning, essentially. Yeah.
Dileep George (56:00.400)
Right. So to finally, to really solve, finally to say that you have solved capture,
Dileep George (56:07.760)
you have to solve the whole problem.
Dileep George (56:08.880)
Yeah. Okay. So how does this kind of the RCN architecture help us to do a better job of that
Dileep George (56:18.560)
kind of thing? Yeah. So as I mentioned, one of the important things was being able to do inference,
Lex Fridman (56:24.960)
being able to dynamically do inference.
Lex Fridman (56:28.640)
Can you clarify what you mean? Because you said like neural networks don't do inference.
Lex Fridman (56:33.040)
Yeah. So what do you mean by inference in this context then?
Dileep George (56:35.840)
So, okay. So in captures, what they do to confuse people is to make these characters crowd together.
Dileep George (56:43.360)
Yes. Okay. And when you make the characters crowd together, what happens is that you will now start
Dileep George (56:48.400)
seeing combinations of characters as some other new character or an existing character. So you
Dileep George (56:53.920)
would put an R and N together. It will start looking like an M. And so locally, there is
Dileep George (57:02.320)
very strong evidence for it being some incorrect character. But globally, the only explanation that
Dileep George (57:11.520)
fits together is something that is different from what you can find locally. Yes. So this is
Dileep George (57:18.240)
inference. You are basically taking local evidence and putting it in the global context and often
Dileep George (57:25.840)
coming to a conclusion locally, which is conflicting with the local information.
Lex Fridman (57:29.920)
So actually, so you mean inference like in the way it's used when you talk about reasoning,
Dileep George (57:36.560)
for example, as opposed to like inference, which is with artificial neural networks,
Dileep George (57:42.240)
which is a single pass to the network. Okay. So like you're basically doing some basic forms of
Dileep George (57:47.840)
reasoning, like integration of like how local things fit into the global picture.
Lex Fridman (57:54.480)
And things like explaining a way coming into this one, because you are explaining that piece
Dileep George (57:59.840)
of evidence as something else, because globally, that's the only thing that makes sense. So now
Dileep George (58:08.160)
you can amortize this inference in a neural network. If you want to do this, you can brute
Dileep George (58:15.600)
force it. You can just show it all combinations of things that you want your reasoning to work over.
Lex Fridman (58:23.120)
And you can just train the help out of that neural network and it will look like it is doing inference
Dileep George (58:30.880)
on the fly, but it is really just doing amortized inference. It is because you have shown it a lot
Dileep George (58:37.680)
of these combinations during training time. So what you want to do is be able to do dynamic
Dileep George (58:43.840)
inference rather than just being able to show all those combinations in the training time.
Lex Fridman (58:48.480)
And that's something we emphasized in the model. What does it mean, dynamic inference? Is that
Dileep George (58:54.080)
that has to do with the feedback thing? Yes. Like what is dynamic? I'm trying to visualize what
Dileep George (59:00.320)
dynamic inference would be in this case. Like what is it doing with the input? It's shown the input
Dileep George (59:05.920)
the first time. Yeah. And is like what's changing over temporally? What's the dynamics of this
Dileep George (59:13.840)
inference process? So you can think of it as you have at the top of the model, the characters that
Dileep George (59:19.840)
you are trained on. They are the causes that you are trying to explain the pixels using the
Dileep George (59:26.720)
characters as the causes. The characters are the things that cause the pixels. Yeah. So there's
Dileep George (59:33.600)
this causality thing. So the reason you mentioned causality, I guess, is because there's a temporal
Dileep George (59:38.960)
aspect to this whole thing. In this particular case, the temporal aspect is not important.
Dileep George (59:43.280)
It is more like when if I turn the character on, the pixels will turn on. Yeah, it will be after
Dileep George (59:50.000)
this a little bit. Okay. So that is causality in the sense of like a logic causality, like
Dileep George (59:55.520)
hence inference. Okay. The dynamics is that even though locally it will look like, okay, this is an
🔗 相关节目