Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics
AI 与机器学习音乐与艺术心理与人性技术与编程政治与社会
📋 章节目录
暂无章节信息
🔑 关键词
dongoingrobotmodelspacelearningreasoningrobotsinterestingproblemskindsscienceuncertaintystateplanninghumancomputerphilosophystuffabstractions
💬 精彩语录
"is harder perception or planning perception? That's why understanding that's why. So what do you think"
感知和计划感知哪个更难？这就是为什么理解就是为什么。那你觉得怎么样
— Leslie Kaelbling （37:43.280）
"it out. Well, yeah. When you try to actually solve the problem with computers, the right answer comes"
它出来了。嗯，是的。当你尝试用计算机实际解决问题时，正确的答案就会出现
— Leslie Kaelbling （17:12.560）
"mere humans. Speaking of which, when you maybe now we take a little step into that philosophy circle."
只是人类。说到这里，当你也许现在我们向那个哲学圈迈出了一小步。
— Leslie Kaelbling （34:10.720）
"MIT. She is recognized for her work in reinforcement learning, planning, robot navigation, and several"
麻省理工学院。她因其在强化学习、规划、机器人导航等领域的工作而受到认可
— Leslie Kaelbling （00:05.360）
"model, which tells me something about the dynamics of the world. If I take it, imagine that I learned"
模型，它告诉我一些有关世界动态的信息。如果我接受它，想象我学到了
— Leslie Kaelbling （36:14.960）
🎙️ 完整对话（607 条）
Lex Fridman (00:00.000)
The following is a conversation with Leslie Kaelbling. She is a roboticist and professor at
以下是与莱斯利·凯尔布林的对话。她是一名机器人专家和教授
Lex Fridman (00:05.360)
MIT. She is recognized for her work in reinforcement learning, planning, robot navigation, and several
麻省理工学院。她因其在强化学习、规划、机器人导航等领域的工作而受到认可
Lex Fridman (00:12.080)
other topics in AI. She won the IJCAI Computers and Thought Award and was the editor in chief
人工智能中的其他主题。荣获IJCAI计算机与思想奖并担任主编
Lex Fridman (00:18.560)
of the prestigious Journal of Machine Learning Research. This conversation is part of the
著名的《机器学习研究杂志》的作者。这段对话是
Lex Fridman (00:24.320)
Artificial Intelligence podcast at MIT and beyond. If you enjoy it, subscribe on YouTube,
麻省理工学院及其他地方的人工智能播客。如果您喜欢，请在 YouTube 上订阅，
Leslie Kaelbling (00:30.400)
iTunes, or simply connect with me on Twitter at Lex Friedman, spelled F R I D.
iTunes，或者直接在 Twitter 上与我联系 Lex Friedman（拼写为 F R I D）。
Lex Fridman (00:36.960)
And now, here's my conversation with Leslie Kaelbling.
现在，这是我与莱斯利·凯尔布林的对话。
Lex Fridman (00:42.800)
What made me get excited about AI, I can say that, is I read Gödel Escher Bach when I was
我可以说，是什么让我对人工智能感到兴奋，是我小时候读过哥德尔·埃舍尔·巴赫的书。
Leslie Kaelbling (00:47.680)
in high school. That was pretty formative for me because it exposed the interestingness of
在高中。这对我来说非常重要，因为它揭示了
Leslie Kaelbling (00:57.200)
primitives and combination and how you can make complex things out of simple parts
基元和组合以及如何用简单的零件制作复杂的东西
Lex Fridman (01:02.320)
and ideas of AI and what kinds of programs might generate intelligent behavior. So...
人工智能的想法以及什么样的程序可能会产生智能行为。所以...
Lex Fridman (01:07.760)
So you first fell in love with AI reasoning logic versus robots?
所以你首先爱上的是人工智能推理逻辑而不是机器人？
Leslie Kaelbling (01:12.720)
Yeah, the robots came because my first job, so I finished an undergraduate degree in philosophy
是的，机器人的出现是因为我的第一份工作，所以我完成了哲学本科学位
Leslie Kaelbling (01:18.160)
at Stanford and was about to finish a master's in computer science. And I got hired at SRI
在斯坦福大学，即将完成计算机科学硕士学位。我被 SRI 录用了
Leslie Kaelbling (01:25.360)
in their AI lab and they were building a robot. It was a kind of a follow on to shaky,
在他们的人工智能实验室里，他们正在建造一个机器人。这是一种对摇摇欲坠的后续，
Lex Fridman (01:30.960)
but all the shaky people were not there anymore. And so my job was to try to get this robot to
但所有颤抖的人都已不复存在。所以我的工作就是尝试让这个机器人
Lex Fridman (01:35.840)
do stuff. And that's really kind of what got me interested in robots.
做事。这确实让我对机器人产生了兴趣。
Lex Fridman (01:39.280)
So maybe taking a small step back to your bachelor's in Stanford in philosophy,
所以也许退一步回到你在斯坦福大学的哲学学士学位，
Leslie Kaelbling (01:44.400)
did master's and PhD in computer science, but the bachelor's in philosophy. So what was that
取得了计算机科学硕士和博士学位，但获得了哲学学士学位。那么那是什么
Lex Fridman (01:49.440)
journey like? What elements of philosophy do you think you bring to your work in computer science?
旅程怎么样？您认为您将哪些哲学元素带入了您的计算机科学工作中？
Lex Fridman (01:55.200)
So it's surprisingly relevant. So the part of the reason that I didn't do a computer
Leslie Kaelbling (01:59.840)
science undergraduate degree was that there wasn't one at Stanford at the time,
Lex Fridman (02:03.440)
but that there's a part of philosophy and in fact, Stanford has a special submajor in
Leslie Kaelbling (02:07.360)
something called now symbolic systems, which is logic, model theory, formal semantics of
Leslie Kaelbling (02:13.280)
natural language. And so that's actually a perfect preparation for work in AI and computer science.
Leslie Kaelbling (02:20.080)
That's kind of interesting. So if you were interested in artificial intelligence,
Lex Fridman (02:26.000)
what kind of majors were people even thinking about taking? What is it in your science?
Lex Fridman (02:31.840)
So besides philosophies, what were you supposed to do if you were fascinated by the idea of creating
Leslie Kaelbling (02:37.120)
intelligence? There weren't enough people who did that for that even to be a conversation.
Leslie Kaelbling (02:41.920)
I mean, I think probably, probably philosophy. I mean, it's interesting in my class,
Leslie Kaelbling (02:48.320)
my graduating class of undergraduate philosophers, probably maybe slightly less than half went on in
Leslie Kaelbling (02:56.800)
computer science, slightly less than half went on in law and like one or two went on in philosophy.
Lex Fridman (03:02.320)
So it was a common kind of connection. Do you think AI researchers have a role
Leslie Kaelbling (03:06.960)
to be part time philosophers or should they stick to the solid science and engineering
Leslie Kaelbling (03:11.440)
without sort of taking the philosophizing tangents? I mean, you work with robots,
Leslie Kaelbling (03:16.240)
you think about what it takes to create intelligent beings. Aren't you the perfect
Leslie Kaelbling (03:21.200)
person to think about the big picture philosophy at all? The parts of philosophy that are closest
Leslie Kaelbling (03:25.920)
to AI, I think, or at least the closest to AI that I think about are stuff like
Leslie Kaelbling (03:29.680)
belief and knowledge and denotation and that kind of stuff. And that's, you know,
Leslie Kaelbling (03:35.040)
it's quite formal. And it's like just one step away from the kinds of computer science work that
Leslie Kaelbling (03:40.800)
we do kind of routinely. I think that there are important questions still about what you can do
Leslie Kaelbling (03:50.160)
with a machine and what you can't and so on. Although at least my personal view is that I'm
Leslie Kaelbling (03:54.560)
completely a materialist. And I don't think that there's any reason why we can't make a robot be
Leslie Kaelbling (04:00.960)
behaviorally indistinguishable from a human. And the question of whether it's
Leslie Kaelbling (04:06.560)
distinguishable internally, whether it's a zombie or not in philosophy terms, I actually don't,
Leslie Kaelbling (04:12.720)
I don't know. And I don't know if I care too much about that.
Leslie Kaelbling (04:15.280)
Right. But there is a philosophical notions. They're mathematical and philosophical because
Lex Fridman (04:20.400)
we don't know so much of how difficult it is. How difficult is the perception problem?
Lex Fridman (04:25.600)
How difficult is the planning problem? How difficult is it to operate in this world successfully?
Leslie Kaelbling (04:30.800)
Because our robots are not currently as successful as human beings in many tasks.
Leslie Kaelbling (04:35.920)
The question about the gap between current robots and human beings borders a little bit
Leslie Kaelbling (04:42.480)
on philosophy. You know, the expanse of knowledge that's required to operate in a human world,
Leslie Kaelbling (04:50.160)
required to operate in this world and the ability to form common sense knowledge, the ability to
Leslie Kaelbling (04:55.520)
reason about uncertainty. Much of the work you've been doing, there's open questions there that,
Lex Fridman (05:02.880)
I don't know, required to activate a certain big picture view.
Leslie Kaelbling (05:07.840)
To me, that doesn't seem like a philosophical gap at all. To me, there is a big technical gap.
Leslie Kaelbling (05:12.960)
There's a huge technical gap, but I don't see any reason why it's more than a technical gap.
Leslie Kaelbling (05:19.360)
Perfect. So, when you mentioned AI, you mentioned SRI, and maybe can you describe to me when you
Leslie Kaelbling (05:28.400)
first fell in love with robotics, with robots or inspired, so you mentioned Flaky or Shaky Flaky,
Lex Fridman (05:38.400)
and what was the robot that first captured your imagination, what's possible?
Leslie Kaelbling (05:42.800)
Right. Well, so the first robot I worked with was Flaky. Shaky was a robot that the SRI
Leslie Kaelbling (05:47.120)
people had built, but by the time, I think when I arrived, it was sitting in a corner of somebody's
Leslie Kaelbling (05:52.960)
office dripping hydraulic fluid into a pan, but it's iconic and really everybody should read the
Leslie Kaelbling (06:00.240)
Shaky Tech Report because it has so many good ideas in it. I mean, they invented ASTAR search
Lex Fridman (06:06.560)
and symbolic planning and learning macro operators. They had low level kind of
Leslie Kaelbling (06:14.240)
configuration space planning for their robot. They had vision. That's the basic ideas of
Leslie Kaelbling (06:19.360)
a ton of things. Can you take a step back? Shaky have arms. What was the job? Shaky was a mobile
Leslie Kaelbling (06:26.240)
robot, but it could push objects, and so it would move things around. With which actuator? With
Leslie Kaelbling (06:32.080)
itself, with its base. Okay, great. And they had painted the baseboards black, so it used vision
Leslie Kaelbling (06:43.600)
to localize itself in a map. It detected objects. It could detect objects that were surprising to
Leslie Kaelbling (06:49.360)
it. It would plan and replan based on what it saw. It reasoned about whether to look and take
Leslie Kaelbling (06:55.520)
pictures. I mean, it really had the basics of so many of the things that we think about now.
Lex Fridman (07:03.280)
How did it represent the space around it? So it had representations at a bunch of different levels
Leslie Kaelbling (07:08.960)
of abstraction. So it had, I think, a kind of an occupancy grid of some sort at the lowest level.
Leslie Kaelbling (07:14.880)
At the high level, it was abstract symbolic kind of rooms and connectivity. So where does flaky
Leslie Kaelbling (07:21.440)
come in? Yeah, okay. So I showed up at SRI and we were building a brand new robot. As I said,
Leslie Kaelbling (07:28.240)
none of the people from the previous project were kind of there or involved anymore. So we were kind
Leslie Kaelbling (07:33.200)
of starting from scratch and my advisor was Stan Rosenstein. He ended up being my thesis advisor
Lex Fridman (07:40.880)
and he was motivated by this idea of situated computation or situated automata. And the idea was
Leslie Kaelbling (07:49.600)
that the tools of logical reasoning were important, but possibly only for the engineers
Leslie Kaelbling (07:58.480)
or designers to use in the analysis of a system, but not necessarily to be manipulated in the head
Leslie Kaelbling (08:04.560)
of the system itself. So I might use logic to prove a theorem about the behavior of my robot,
Leslie Kaelbling (08:10.480)
even if the robot's not using logic in its head to prove theorems. So that was kind of the
Leslie Kaelbling (08:14.400)
distinction. And so the idea was to kind of use those principles to make a robot do stuff. But
Leslie Kaelbling (08:23.440)
a lot of the basic things we had to kind of learn for ourselves because I had zero background in
Leslie Kaelbling (08:29.600)
robotics. I didn't know anything about control. I didn't know anything about sensors. So we
Leslie Kaelbling (08:33.600)
reinvented a lot of wheels on the way to getting that robot to do stuff. Do you think that was
Leslie Kaelbling (08:37.280)
an advantage or a hindrance? Oh no, I mean, I'm big in favor of wheel reinvention actually. I mean,
Leslie Kaelbling (08:44.800)
I think you learn a lot by doing it. It's important though to eventually have the pointers
Leslie Kaelbling (08:49.680)
to, so that you can see what's really going on. But I think you can appreciate much better the
Leslie Kaelbling (08:56.640)
good solutions once you've messed around a little bit on your own and found a bad one.
Leslie Kaelbling (09:00.400)
Yeah. I think you mentioned reinventing reinforcement learning and referring to
Leslie Kaelbling (09:04.880)
rewards as pleasures, pleasure. Yeah. Or I think, which I think is a nice name for it.
Leslie Kaelbling (09:11.440)
Yeah. It's more fun almost. Do you think you could tell the history of AI machine learning
Lex Fridman (09:18.960)
reinforcement learning and how you think about it from the fifties to now?
Leslie Kaelbling (09:23.600)
One thing is that it's oscillates, right? So things become fashionable and then they go out
Lex Fridman (09:29.360)
and then something else becomes cool and that goes out and so on. And I think there's, so there's
Leslie Kaelbling (09:33.680)
some interesting sociological process that actually drives a lot of what's going on.
Leslie Kaelbling (09:38.880)
Early days was kind of cybernetics and control, right? And the idea that of homeostasis,
Leslie Kaelbling (09:46.240)
right? People have made these robots that could, I don't know, try to plug into the wall when they
Leslie Kaelbling (09:51.680)
needed power and then come loose and roll around and do stuff. And then I think over time, the
Leslie Kaelbling (09:59.200)
thought, well, that was inspiring, but people said, no, no, no, we want to get maybe closer to what
Leslie Kaelbling (10:03.200)
feels like real intelligence or human intelligence. And then maybe the expert systems people tried
Leslie Kaelbling (10:10.160)
to do that, but maybe a little too superficially, right? So, oh, we get the surface understanding of
Lex Fridman (10:20.240)
what intelligence is like, because I understand how a steel mill works and I can try to explain
Leslie Kaelbling (10:24.960)
it to you and you can write it down in logic and then we can make a computer and for that.
Lex Fridman (10:29.200)
And then that didn't work out. But what's interesting, I think, is when a thing starts to not
Leslie Kaelbling (10:36.400)
be working very well, it's not only do we change methods, we change problems, right? So it's not
Leslie Kaelbling (10:43.200)
like we have better ways of doing the problem of the expert systems people were trying to do. We
Leslie Kaelbling (10:47.040)
have no ways of trying to do that problem. Oh, yeah, no, I think maybe a few, but we kind of
Leslie Kaelbling (10:54.880)
give up on that problem and we switched to a different problem and we worked that for a while
Lex Fridman (11:00.720)
and we make progress. As a broad community. As a community, yeah. And there's a lot of people who
Leslie Kaelbling (11:04.320)
would argue, you don't give up on the problem, it's just you decrease the number of people working
Leslie Kaelbling (11:09.520)
on it. You almost kind of like put it on the shelf, say, we'll come back to this 20 years later.
Leslie Kaelbling (11:13.920)
Yeah, I think that's right. Or you might decide that it's malformed. Like you might say,
Leslie Kaelbling (11:21.600)
it's wrong to just try to make something that does superficial symbolic reasoning
Leslie Kaelbling (11:26.160)
behave like a doctor. You can't do that until you've had the sensory motor experience of being
Leslie Kaelbling (11:33.200)
a doctor or something. So there's arguments that say that that problem was not well formed. Or it
Leslie Kaelbling (11:38.320)
could be that it is well formed, but we just weren't approaching it well. So you mentioned
Leslie Kaelbling (11:43.760)
that your favorite part of logic and symbolic systems is that they give short names for large
Leslie Kaelbling (11:48.800)
sets. So there is some use to this. They use symbolic reasoning. So looking at expert systems
Lex Fridman (11:56.960)
and symbolic computing, what do you think are the roadblocks that were hit in the 80s and 90s?
Leslie Kaelbling (12:01.680)
Ah, okay. So right. So the fact that I'm not a fan of expert systems doesn't mean that I'm not a
Leslie Kaelbling (12:07.920)
fan of some kinds of symbolic reasoning, right? So let's see, roadblocks. Well, the main road
Leslie Kaelbling (12:15.520)
block, I think, was that the idea that humans could articulate their knowledge effectively
Leslie Kaelbling (12:22.080)
into some kind of logical statements.
Lex Fridman (12:26.240)
So it's not just the cost, the effort, but really just the capability of doing it.
Leslie Kaelbling (12:31.280)
Right. Because we're all experts in vision, right? But totally don't have introspective
Leslie Kaelbling (12:36.720)
access into how we do that. Right. And it's true that, I mean, I think the idea was, well,
Leslie Kaelbling (12:44.960)
of course, even people then would know, of course, I wouldn't ask you to please write
Leslie Kaelbling (12:48.240)
down the rules that you use for recognizing a water bottle. That's crazy. And everyone
Leslie Kaelbling (12:52.800)
understood that. But we might ask you to please write down the rules you use for deciding,
Leslie Kaelbling (12:58.640)
I don't know, what tie to put on or how to set up a microphone or something like that.
Lex Fridman (13:04.640)
But even those things, I think people maybe, I think what they found, I'm not sure about
Leslie Kaelbling (13:10.880)
this, but I think what they found was that the so called experts could give explanations
Leslie Kaelbling (13:16.000)
that sort of post hoc explanations for how and why they did things, but they weren't
Leslie Kaelbling (13:20.080)
necessarily very good. And then they depended on maybe some kinds of perceptual things,
Leslie Kaelbling (13:28.960)
which again, they couldn't really define very well. So I think fundamentally, I think the
Leslie Kaelbling (13:35.840)
underlying problem with that was the assumption that people could articulate how and why they
Leslie Kaelbling (13:40.960)
make their decisions. Right. So it's almost encoding the knowledge
Leslie Kaelbling (13:45.680)
from converting from expert to something that a machine could understand and reason with.
Leslie Kaelbling (13:51.440)
No, no, no, no, not even just encoding, but getting it out of you.
Leslie Kaelbling (13:56.320)
Right. Not, not, not writing it. I mean, yes, hard also to write it down for the computer,
Lex Fridman (14:02.240)
but I don't think that people can produce it. You can tell me a story about why you do stuff,
Lex Fridman (14:08.320)
but I'm not so sure that's the why. Great. So there are still on the
Leslie Kaelbling (14:14.400)
hierarchical planning side, places where symbolic reasoning is very useful. So as you've talked
Leslie Kaelbling (14:24.560)
about, so where's the gap? Yeah. Okay, good. So saying that humans can't provide a description
Leslie Kaelbling (14:34.960)
of their reasoning processes. That's okay. Fine. But that doesn't mean that it's not good to do
Leslie Kaelbling (14:41.040)
reasoning of various styles inside a computer. Those are just two orthogonal points. So then
Leslie Kaelbling (14:47.120)
the question is what kind of reasoning should you do inside a computer? Right.
Lex Fridman (14:52.240)
And the answer is, I think you need to do all different kinds of reasoning inside a computer,
Leslie Kaelbling (14:56.240)
depending on what kinds of problems you face. I guess the question is what kind of things can you
Leslie Kaelbling (15:02.400)
encode symbolically so you can reason about? I think the idea about, and even symbolic,
Leslie Kaelbling (15:12.480)
I don't even like that terminology because I don't know what it means technically and formally.
Leslie Kaelbling (15:17.760)
I do believe in abstractions. So abstractions are critical, right? You cannot reason at completely
Leslie Kaelbling (15:24.160)
fine grain about everything in your life, right? You can't make a plan at the level of images and
Leslie Kaelbling (15:29.520)
torques for getting a PhD. So you have to reduce the size of the state space and you have to reduce
Leslie Kaelbling (15:36.960)
the horizon if you're going to reason about getting a PhD or even buying the ingredients to
Lex Fridman (15:42.160)
make dinner. And so how can you reduce the spaces and the horizon of the reasoning you have to do?
Lex Fridman (15:49.280)
And the answer is abstraction, spatial abstraction, temporal abstraction. I think abstraction along
Leslie Kaelbling (15:53.760)
the lines of goals is also interesting, like you might, well, abstraction and decomposition. Goals
Leslie Kaelbling (16:00.480)
is maybe more of a decomposition thing. So I think that's where these kinds of, if you want to call
Leslie Kaelbling (16:05.600)
it symbolic or discrete models come in. You talk about a room of your house instead of your pose.
Leslie Kaelbling (16:12.800)
You talk about doing something during the afternoon instead of at 2.54. And you do that because it
Lex Fridman (16:20.800)
and you do that because it makes your reasoning problem easier. And also because
Leslie Kaelbling (16:27.120)
you have, you don't have enough information to reason in high fidelity about your pose of your
Lex Fridman (16:34.800)
elbow at 2.35 this afternoon anyway. Right. When you're trying to get a PhD.
Leslie Kaelbling (16:39.520)
Or when you're doing anything really. Yeah. Okay.
Leslie Kaelbling (16:42.720)
Except for at that moment, at that moment, you do have to reason about the pose of your elbow,
Leslie Kaelbling (16:46.080)
maybe, but then you, maybe you do that in some continuous joint space kind of model.
Lex Fridman (16:50.000)
And so again, I, my biggest point about all of this is that there should be the dogma is not
Leslie Kaelbling (16:58.640)
the thing, right? We shouldn't, it shouldn't be that I'm in favor against symbolic reasoning
Lex Fridman (17:02.880)
and you're in favor against neural networks. It should be that just, just computer science
Leslie Kaelbling (17:08.800)
tells us what the right answer to all these questions is. If we were smart enough to figure
Leslie Kaelbling (17:12.560)
it out. Well, yeah. When you try to actually solve the problem with computers, the right answer comes
Leslie Kaelbling (17:17.200)
out. But you mentioned abstractions. I mean, neural networks form abstractions or rather
Leslie Kaelbling (17:24.480)
there's, there's automated ways to form abstractions and there's expert driven ways to
Leslie Kaelbling (17:28.960)
form abstractions and expert human driven ways. And humans just seem to be way better at forming
Leslie Kaelbling (17:35.280)
abstractions currently and certain problems. So when you're referring to 2.45 PM versus afternoon,
Lex Fridman (17:44.000)
how do we construct that taxonomy? Is there any room for automated construction of such
Leslie Kaelbling (17:50.080)
abstractions? Oh, I think eventually, yeah. I mean, I think when we get to be better
Lex Fridman (17:55.280)
and machine learning engineers, we'll build algorithms that build awesome abstractions.
Leslie Kaelbling (18:01.120)
That are useful in this kind of way that you're describing. Yeah. So let's then step from
Leslie Kaelbling (18:05.760)
the, the abstraction discussion and let's talk about POMM MDPs. Partially observable
Lex Fridman (18:14.800)
Markov decision processes. So uncertainty. So first, what are Markov decision processes?
Lex Fridman (18:20.080)
What are Markov decision processes? And maybe how much of our world can be models and MDPs? How
Leslie Kaelbling (18:26.320)
much, when you wake up in the morning and you're making breakfast, how do you, do you think of
Leslie Kaelbling (18:30.160)
yourself as an MDP? So how do you think about MDPs and how they relate to our world? Well, so
Leslie Kaelbling (18:36.640)
there's a stance question, right? So a stance is a position that I take with respect to a problem.
Lex Fridman (18:42.240)
So I, as a researcher or a person who designs systems, can decide to make a model of the world
Leslie Kaelbling (18:50.720)
around me in some terms. So I take this messy world and I say, I'm going to treat it as if it
Leslie Kaelbling (18:57.600)
were a problem of this formal kind, and then I can apply solution concepts or algorithms or whatever
Leslie Kaelbling (19:02.960)
to solve that formal thing, right? So of course the world is not anything. It's not an MDP or a
Leslie Kaelbling (19:07.920)
POMM DP. I don't know what it is, but I can model aspects of it in some way or some other way.
Lex Fridman (19:12.880)
And when I model some aspect of it in a certain way, that gives me some set of algorithms I can
Leslie Kaelbling (19:17.600)
use. You can model the world in all kinds of ways. Some have, some are, some are, some are
Leslie Kaelbling (19:26.160)
more accepting of uncertainty, more easily modeling uncertainty of the world. Some really force the
Leslie Kaelbling (19:33.360)
world to be deterministic. And so certainly MDPs model the uncertainty of the world. Yes. Model
Leslie Kaelbling (19:42.000)
some uncertainty. They model not present state uncertainty, but they model uncertainty in the
Leslie Kaelbling (19:47.520)
way the future will unfold. Right. So what are Markov decision processes? So Markov decision
Leslie Kaelbling (19:54.560)
process is a model. It's a kind of a model that you could make that says, I know completely the
Leslie Kaelbling (19:59.760)
current state of my system. And what it means to be a state is that I, that all the, I have all
Leslie Kaelbling (1:00:00.640)
it's a combination of learning and not learning. And what should that combination be? And what's
Leslie Kaelbling (1:00:05.760)
the stuff we build in? So to me, that's the most compelling question. And when you say engineer
Lex Fridman (1:00:10.080)
robots, you mean engineering systems that work in the real world? Is that, that's the emphasis?
Lex Fridman (1:00:16.560)
Okay. Last question. Which robots or robot is your favorite from science fiction?
Lex Fridman (1:00:24.480)
So you can go with Star Wars or RTD2, or you can go with more modern,
Leslie Kaelbling (1:00:32.240)
maybe Hal from... I don't think I have a favorite robot from science fiction.
Leslie Kaelbling (1:00:38.080)
This is, this is back to, you like to make robots work in the real world here, not, not in...
Leslie Kaelbling (1:00:45.520)
I mean, I love the process and I care more about the process.
Lex Fridman (1:00:50.000)
The engineering process.
Leslie Kaelbling (1:00:51.600)
Yeah. I mean, I do research because it's fun, not because I care about what we produce.
Lex Fridman (1:00:57.520)
Well, that's a, that's a beautiful note actually. And Leslie,
Leslie Kaelbling (1:01:00.640)
thank you so much for talking today.
Lex Fridman (1:01:02.000)
Sure. It's been fun.
Leslie Kaelbling (20:05.760)
the information right now that will let me make predictions about the future as well as I can.
Lex Fridman (20:11.120)
So that remembering anything about my history wouldn't make my predictions any better.
Leslie Kaelbling (20:17.760)
And, but then it also says that then I can take some actions that might change the state of the
Leslie Kaelbling (20:23.280)
world. And that I don't have a deterministic model of those changes. I have a probabilistic
Leslie Kaelbling (20:28.400)
model of how the world might change. It's a, it's a useful model for some kinds of systems.
Leslie Kaelbling (20:34.160)
I think it's a, I mean, it's certainly not a good model for most problems, I think, because for most
Leslie Kaelbling (20:42.480)
problems you don't actually know the state. For most problems you, it's partially observed. So
Leslie Kaelbling (20:48.720)
that's now a different problem class. So, okay. That's where the POMDPs, the part that we observe
Leslie Kaelbling (20:55.600)
with the Markov decision processes step in. So how do they address the fact that you can't
Leslie Kaelbling (21:01.360)
observe most incomplete information about most of the world around you? Right. So now the idea is
Leslie Kaelbling (21:07.760)
we still kind of postulate that there exists a state. We think that there is some information
Leslie Kaelbling (21:12.880)
about the world out there such that if we knew that we could make good predictions, but we don't
Leslie Kaelbling (21:18.000)
know the state. And so then we have to think about how, but we do get observations. Maybe I get
Leslie Kaelbling (21:23.680)
images or I hear things or I feel things, and those might be local or noisy. And so therefore
Leslie Kaelbling (21:29.680)
they don't tell me everything about what's going on. And then I have to reason about given the
Leslie Kaelbling (21:34.720)
history of actions I've taken and observations I've gotten, what do I think is going on in the
Leslie Kaelbling (21:39.920)
world? And then given my own kind of uncertainty about what's going on in the world, I can decide
Lex Fridman (21:43.760)
what actions to take. And so how difficult is this problem of planning under uncertainty in your
Leslie Kaelbling (21:50.080)
view and your long experience of modeling the world, trying to deal with this uncertainty in
Leslie Kaelbling (21:57.840)
especially in real world systems? Optimal planning for even discrete POMDPs can be undecidable
Leslie Kaelbling (22:05.040)
depending on how you set it up. And so lots of people say, I don't use POMDPs because they are
Leslie Kaelbling (22:12.960)
intractable. And I think that that's kind of a very funny thing to say because the problem you
Leslie Kaelbling (22:19.600)
have to solve is the problem you have to solve. So if the problem you have to solve is intractable,
Leslie Kaelbling (22:24.320)
that's what makes us AI people, right? So we solve, we understand that the problem we're
Leslie Kaelbling (22:28.720)
solving is wildly intractable that we can't, we will never be able to solve it optimally,
Leslie Kaelbling (22:34.320)
at least I don't. Yeah, right. So later we can come back to an idea about bounded optimality
Lex Fridman (22:41.360)
and something. But anyway, we can't come up with optimal solutions to these problems.
Lex Fridman (22:45.520)
So we have to make approximations, approximations in modeling, approximations in the solution
Leslie Kaelbling (22:50.640)
algorithms and so on. And so I don't have a problem with saying, yeah, my problem actually,
Leslie Kaelbling (22:56.960)
it is POMDP in continuous space with continuous observations. And it's so computationally complex,
Leslie Kaelbling (23:02.480)
I can't even think about it's, you know, big O whatever. But that doesn't prevent me from,
Leslie Kaelbling (23:09.600)
it helps me, gives me some clarity to think about it that way and to then take steps to
Leslie Kaelbling (23:16.160)
make approximation after approximation to get down to something that's like computable
Leslie Kaelbling (23:20.880)
in some reasonable time. When you think about optimality, the community broadly has shifted on
Leslie Kaelbling (23:26.720)
that, I think a little bit in how much they value the idea of optimality, of chasing an optimal
Leslie Kaelbling (23:34.880)
solution. How has your views of chasing an optimal solution changed over the years when
Leslie Kaelbling (23:40.960)
you work with robots? That's interesting. I think we have a little bit of a methodological crisis
Leslie Kaelbling (23:48.880)
actually from the theoretical side. I mean, I do think that theory is important and that right now
Leslie Kaelbling (23:53.520)
we're not doing much of it. So there's lots of empirical hacking around and training this and
Leslie Kaelbling (24:00.080)
doing that and reporting numbers, but is it good? Is it bad? We don't know. It's very hard to say
Leslie Kaelbling (24:05.120)
things. And if you look at like computer science theory, so people talked for a while, everyone was
Leslie Kaelbling (24:16.400)
about solving problems optimally or completely. And then there were interesting relaxations. So
Leslie Kaelbling (24:22.320)
people look at, oh, are there regret bounds or can I do some kind of approximation? Can I prove
Leslie Kaelbling (24:30.640)
something that I can approximately solve this problem or that I get closer to the solution as
Leslie Kaelbling (24:34.880)
I spend more time and so on? What's interesting I think is that we don't have good approximate
Leslie Kaelbling (24:42.560)
solution concepts for very difficult problems. I like to say that I'm interested in doing a very
Leslie Kaelbling (24:51.520)
bad job of very big problems. Right. So very bad job, very big problems. I like to do that,
Lex Fridman (25:00.400)
but I wish I could say something. I wish I had a, I don't know, some kind of a formal solution
Leslie Kaelbling (25:09.120)
concept that I could use to say, oh, this algorithm actually, it gives me something.
Leslie Kaelbling (25:16.240)
Like I know what I'm going to get. I can do something other than just run it and get out.
Lex Fridman (25:19.840)
So that, that notion is still somewhere deeply compelling to you. The notion that you can say,
Leslie Kaelbling (25:27.520)
you can drop thing on the table says this, you can expect this, this algorithm will
Leslie Kaelbling (25:32.400)
give me some good results. I hope there's, I hope science will, I mean,
Leslie Kaelbling (25:37.280)
there's engineering and there's science. I think that they're not exactly the same.
Lex Fridman (25:42.320)
And I think right now we're making huge engineering, like leaps and bounds. So the
Leslie Kaelbling (25:47.040)
engineering is running away ahead of the science, which is cool. And often how it goes, right? So
Leslie Kaelbling (25:52.240)
we're making things and nobody knows how and why they work roughly, but we need to turn that into
Leslie Kaelbling (25:59.680)
science. There's some form. It's a, yeah, there's some room for formalizing. We need to know what
Leslie Kaelbling (26:05.440)
the principles are. Why does this work? Why does that not work? I mean, for a while, people built
Leslie Kaelbling (26:09.840)
bridges by trying, but now we can often predict whether it's going to work or not without building
Leslie Kaelbling (26:14.480)
it. Can we do that for learning systems or for robots? So your hope is from a materialistic
Leslie Kaelbling (26:20.640)
perspective that intelligence, artificial intelligence systems, robots are just fancier
Leslie Kaelbling (26:27.600)
bridges. Belief space. What's the difference between belief space and state space? So you
Lex Fridman (26:33.040)
mentioned MDPs, FOMDPs, reasoning about, you sense the world, there's a state.
Leslie Kaelbling (26:39.840)
Uh, what, what's this belief space idea? That sounds so good.
Leslie Kaelbling (26:44.400)
It sounds good. So belief space, that is instead of thinking about what's the state of the world
Lex Fridman (26:51.600)
and trying to control that as a robot, I think about what is the space of beliefs that I could
Leslie Kaelbling (26:58.880)
have about the world. What's, if I think of a belief as a probability distribution of our ways
Leslie Kaelbling (27:03.520)
the world could be, a belief state is a distribution. And then my control problem, if I'm reasoning
Leslie Kaelbling (27:10.080)
about how to move through a world I'm uncertain about, my control problem is actually the problem
Leslie Kaelbling (27:16.160)
of controlling my beliefs. So I think about taking actions, not just what effect they'll have on the
Leslie Kaelbling (27:21.360)
world outside, but what effect they'll have on my own understanding of the world outside. And so
Leslie Kaelbling (27:26.080)
that might compel me to ask a question or look somewhere to gather information, which may not
Leslie Kaelbling (27:32.800)
really change the world state, but it changes my own belief about the world. That's a powerful way
Leslie Kaelbling (27:38.400)
to, to empower the agent, to reason about the world, to explore the world. So what kind of
Lex Fridman (27:46.320)
problems does it allow you to solve to, to consider belief space versus just state space?
Leslie Kaelbling (27:52.400)
Well, any problem that requires deliberate information gathering, right? So if in some
Leslie Kaelbling (27:58.400)
problems like chess, there's no uncertainty, or maybe there's uncertainty about the opponent,
Leslie Kaelbling (28:05.040)
there's no uncertainty about the state. And some problems, there's uncertainty,
Lex Fridman (28:10.320)
but you gather information as you go, right? You might say, Oh, I'm driving my autonomous car down
Leslie Kaelbling (28:16.080)
the road and it doesn't know perfectly where it is, but the light hours are all going all the time.
Lex Fridman (28:20.560)
So I don't have to think about whether to gather information. But if you're a human driving down
Leslie Kaelbling (28:25.360)
the road, you sometimes look over your shoulder to see what's going on behind you in the lane.
Lex Fridman (28:31.360)
And you have to decide whether you should do that now. And you have to trade off the fact that
Leslie Kaelbling (28:37.840)
you're not seeing in front of you and you're looking behind you and how valuable is that
Leslie Kaelbling (28:41.520)
information and so on. And so to make choices about information gathering, you have to reasonably
Leslie Kaelbling (28:47.200)
space. Also, I mean, also to just take into account your own uncertainty before trying to
Leslie Kaelbling (28:56.800)
do things. So you might say, if I understand where I'm standing relative to the door jam,
Leslie Kaelbling (29:05.360)
pretty accurately, then it's okay for me to go through the door. But if I'm really
Lex Fridman (29:08.640)
not sure where the door is, then it might be better to not do that right now.
Leslie Kaelbling (29:12.960)
The degree of your uncertainty about the world is actually part of the thing you're trying to
Leslie Kaelbling (29:17.760)
optimize in forming the plan, right? So this idea of a long horizon of planning for a PhD or just
Leslie Kaelbling (29:25.760)
even how to get out of the house or how to make breakfast. You show this presentation of the WTF,
Leslie Kaelbling (29:31.920)
where's the fork of robot looking at a sink. And can you describe how we plan in this world
Leslie Kaelbling (29:40.640)
of this idea of hierarchical planning we've mentioned? So yeah, how can a robot hope to
Lex Fridman (29:47.360)
plan about something with such a long horizon where the goal is quite far away?
Leslie Kaelbling (29:54.480)
People since probably reasoning began have thought about hierarchical reasoning,
Leslie Kaelbling (29:59.840)
the temporal hierarchy in particular. Well, there's spatial hierarchy, but let's talk
Leslie Kaelbling (30:03.040)
about temporal hierarchy. So you might say, oh, I have this long execution I have to do,
Lex Fridman (30:08.960)
but I can divide it into some segments abstractly, right? So maybe you have to get out of the house,
Leslie Kaelbling (30:15.360)
I have to get in the car, I have to drive and so on. And so you can plan if you can build
Leslie Kaelbling (30:22.720)
abstractions. So this we started out by talking about abstractions. And we're back to that now,
Leslie Kaelbling (30:26.960)
if you can build abstractions in your state space, and abstractions sort of temporal abstractions,
Leslie Kaelbling (30:34.560)
then you can make plans at a high level. And you can say, I'm going to go to town and then I'll
Leslie Kaelbling (30:39.760)
have to get gas and then I can go here and I can do this other thing. And you can reason about the
Leslie Kaelbling (30:43.520)
dependencies and constraints among these actions, again, without thinking about the complete
Leslie Kaelbling (30:50.080)
details. What we do in our hierarchical planning work is then say, all right, I make a plan at a
Leslie Kaelbling (30:56.800)
high level of abstraction, I have to have some reason to think that it's feasible without working
Leslie Kaelbling (31:03.760)
it out in complete detail. And that's actually the interesting step. I always like to talk about
Leslie Kaelbling (31:08.720)
walking through an airport, like you can plan to go to New York and arrive at the airport, and then
Leslie Kaelbling (31:15.120)
find yourself an office building later. You can't even tell me in advance what your plan is for
Leslie Kaelbling (31:20.080)
walking through the airport, partly because you're too lazy to think about it, maybe, but partly
Leslie Kaelbling (31:24.960)
also because you just don't have the information, you don't know what gate you're landing in, or
Lex Fridman (31:28.800)
what people are going to be in front of you or anything. So there's no point in planning in
Leslie Kaelbling (31:34.080)
detail, but you have to have, you have to make a leap of faith that you can figure it out once you
Leslie Kaelbling (31:40.640)
get there. And it's really interesting to me how you arrive at that. How do you, so you have learned
Leslie Kaelbling (31:50.000)
over your lifetime to be able to make some kinds of predictions about how hard it is to achieve some
Leslie Kaelbling (31:54.720)
kinds of sub goals. And that's critical. Like you would never plan to fly somewhere if you couldn't,
Leslie Kaelbling (32:00.720)
didn't have a model of how hard it was to do some of the intermediate steps. So one of the things
Leslie Kaelbling (32:04.800)
we're thinking about now is how do you do this kind of very aggressive generalization to situations
Leslie Kaelbling (32:12.480)
that you haven't been in and so on to predict how long will it take to walk through the Kuala Lumpur
Leslie Kaelbling (32:16.880)
airport. Like you could give me an estimate and it wouldn't be crazy. And you have to have an
Leslie Kaelbling (32:22.480)
estimate of that in order to make plans that involve walking through the Kuala Lumpur airport,
Leslie Kaelbling (32:27.280)
even if you don't need to know it in detail. So I'm really interested in these kinds of abstract
Leslie Kaelbling (32:32.480)
models and how do we acquire them. But once we have them, we can use them to do hierarchical
Leslie Kaelbling (32:37.280)
reasoning, which is, I think is very important. Yeah. There's this notion of goal regression and
Leslie Kaelbling (32:43.280)
preimage backchaining, this idea of starting at the goal and just forming these big clouds of
Leslie Kaelbling (32:50.080)
states. I mean, it's almost like saying to the airport, you know, once you show up to the airport
Leslie Kaelbling (33:01.920)
that you're like a few steps away from the goal. So like thinking of it this way, it's kind of
Leslie Kaelbling (33:08.160)
interesting. I don't know if you have sort of further comments on that of starting at the goal.
Leslie Kaelbling (33:14.480)
Yeah. I mean, it's interesting that Simon, Herb Simon back in the early days of AI talked a lot
Leslie Kaelbling (33:22.000)
about means ends reasoning and reasoning back from the goal. There's a kind of an intuition that
Leslie Kaelbling (33:26.480)
people have that the number of that state space is big. The number of actions you could take is
Leslie Kaelbling (33:34.480)
really big. So if you say, here I sit and I want to search forward from where I am, what are all
Leslie Kaelbling (33:39.120)
the things I could do? That's just overwhelming. If you say, if you can reason at this other level
Lex Fridman (33:44.720)
and say, here's what I'm hoping to achieve, what could I do to make that true? That somehow the
Leslie Kaelbling (33:49.520)
branching is smaller. Now what's interesting is that like in the AI planning community,
Leslie Kaelbling (33:54.880)
that hasn't worked out in the class of problems that they look at and the methods that they tend
Leslie Kaelbling (33:59.120)
to use. It hasn't turned out that it's better to go backward. It's still kind of my intuition that
Leslie Kaelbling (34:04.560)
it is, but I can't prove that to you right now. Right. I share your intuition, at least for us
Leslie Kaelbling (34:10.720)
mere humans. Speaking of which, when you maybe now we take a little step into that philosophy circle.
Lex Fridman (34:22.400)
How hard would it, when you think about human life, you give those examples often. How hard do
Leslie Kaelbling (34:28.080)
you think it is to formulate human life as a planning problem or aspects of human life? So
Leslie Kaelbling (34:33.360)
when you look at robots, you're often trying to think about object manipulation,
Leslie Kaelbling (34:38.640)
tasks about moving a thing. When you take a slight step outside the room, let the robot
Leslie Kaelbling (34:46.240)
leave and go get lunch, or maybe try to pursue more fuzzy goals. How hard do you think is that
Leslie Kaelbling (34:54.480)
problem? If you were to try to maybe put another way, try to formulate human life as a planning
Leslie Kaelbling (35:00.800)
problem. Well, that would be a mistake. I mean, it's not all a planning problem, right? I think
Leslie Kaelbling (35:05.760)
it's really, really important that we understand that you have to put together pieces and parts
Leslie Kaelbling (35:11.920)
that have different styles of reasoning and representation and learning. I think it seems
Leslie Kaelbling (35:18.640)
probably clear to anybody that it can't all be this or all be that. Brains aren't all like this
Leslie Kaelbling (35:25.680)
or all like that, right? They have different pieces and parts and substructure and so on.
Lex Fridman (35:30.160)
So I don't think that there's any good reason to think that there's going to be like one true
Leslie Kaelbling (35:34.320)
algorithmic thing that's going to do the whole job. So it's a bunch of pieces together designed
Leslie Kaelbling (35:40.720)
to solve a bunch of specific problems. Or maybe styles of problems. I mean, there's probably some
Leslie Kaelbling (35:49.120)
reasoning that needs to go on in image space. I think, again, there's this model based versus
Leslie Kaelbling (35:57.360)
model free idea, right? So in reinforcement learning, people talk about, oh, should I learn,
Leslie Kaelbling (36:02.480)
I could learn a policy, just straight up a way of behaving. I could learn it's popular
Leslie Kaelbling (36:08.400)
on a value function. That's some kind of weird intermediate ground. Or I could learn a transition
Leslie Kaelbling (36:14.960)
model, which tells me something about the dynamics of the world. If I take it, imagine that I learned
Leslie Kaelbling (36:20.240)
a transition model and I couple it with a planner and I draw a box around that, I have a policy
Leslie Kaelbling (36:25.600)
again. It's just stored a different way, right? But it's just as much of a policy as the other
Leslie Kaelbling (36:32.800)
policy. It's just I've made, I think the way I see it is it's a time space trade off in computation,
Leslie Kaelbling (36:40.160)
right? A more overt policy representation. Maybe it takes more space, but maybe I can
Leslie Kaelbling (36:46.160)
compute quickly what action I should take. On the other hand, maybe a very compact model of
Leslie Kaelbling (36:51.200)
the world dynamics plus a planner lets me compute what action to take to just more slowly. There's
Leslie Kaelbling (36:57.120)
no, I don't, I mean, I don't think there's no argument to be had. It's just like a question of
Lex Fridman (37:02.320)
what form of computation is best for us for the various sub problems. Right. So, and, and so like
Leslie Kaelbling (37:10.320)
learning to do algebra manipulations for some reason is, I mean, that's probably gonna want
Leslie Kaelbling (37:16.000)
naturally a sort of a different representation than writing a unicycle at the time constraints
Leslie Kaelbling (37:21.280)
on the unicycle are serious. The space is maybe smaller. I don't know, but so I could be the more
Leslie Kaelbling (37:27.760)
human size of falling in love, having a relationship that might be another, another style of how to
Leslie Kaelbling (37:36.000)
model that. Yeah. Let's first solve the algebra and the object manipulation. What do you think
Leslie Kaelbling (37:43.280)
is harder perception or planning perception? That's why understanding that's why. So what do you think
Leslie Kaelbling (37:52.160)
is so hard about perception by understanding the world around you? Well, I mean, I think the big
Leslie Kaelbling (37:56.480)
question is representational. Hugely the question is representation. So perception has made great
Leslie Kaelbling (38:08.160)
strides lately, right? And we can classify images and we can play certain kinds of games and predict
Lex Fridman (38:15.360)
how to steer the car and all this sort of stuff. Um, I don't think we have a very good idea of
Lex Fridman (38:24.800)
what perception should deliver, right? So if you, if you believe in modularity, okay, there's,
Leslie Kaelbling (38:29.760)
there's a very strong view which says we shouldn't build in any modularity. We should make a giant
Leslie Kaelbling (38:38.000)
gigantic neural network, train it end to end to do the thing. And that's the best way forward.
Lex Fridman (38:44.320)
And it's hard to argue with that except on a sample complexity basis, right? So you might say,
Leslie Kaelbling (38:51.440)
Oh, well if I want to do end to end reinforcement learning on this giant, giant neural network,
Leslie Kaelbling (38:55.280)
it's going to take a lot of data and a lot of like broken robots and stuff. So then the only answer
Leslie Kaelbling (39:05.520)
is to say, okay, we have to build something in, build in some structure or some bias. We know
Leslie Kaelbling (39:11.760)
from theory of machine learning, the only way to cut down the sample complexity is to kind of cut
Leslie Kaelbling (39:15.760)
down, somehow cut down the hypothesis space. You can do that by building in bias. There's all kinds
Leslie Kaelbling (39:22.480)
of reasons to think that nature built bias into humans. Um, convolution is a bias, right? It's a
Leslie Kaelbling (39:30.640)
very strong bias and it's a very critical bias. So my own view is that we should look for more
Leslie Kaelbling (39:37.520)
things that are like convolution, but the address other aspects of reasoning, right? So convolution
Leslie Kaelbling (39:42.880)
helps us a lot with a certain kind of spatial reasoning. That's quite close to the imaging.
Leslie Kaelbling (39:48.320)
I think there's other ideas like that. Maybe some amount of forward search, maybe some notions of
Leslie Kaelbling (39:56.880)
abstraction, maybe the notion that objects exist. Actually, I think that's pretty important. And a
Leslie Kaelbling (40:02.080)
lot of people won't give you that to start with. Right? So almost like a convolution in the, uh,
Leslie Kaelbling (40:08.960)
uh, in the object, semantic object space or some kind of, some kind of ideas in there.
Leslie Kaelbling (40:13.840)
That's right. And people are starting like the graph, graph convolutions are an idea that are
Leslie Kaelbling (40:17.760)
related to relation, relational representations. And so, so I think there are, so you, I've come
Leslie Kaelbling (40:25.840)
I've come far field from perception, but I think, um, I think the thing that's going to make
Leslie Kaelbling (40:30.720)
perception that kind of the next step is actually understanding better what it should produce.
Leslie Kaelbling (40:36.800)
Right? So what are we going to do with the output of it? Right? It's fine when what we're going to
Leslie Kaelbling (40:40.640)
do with the output is steer. It's less clear when we're just trying to make a one integrated
Leslie Kaelbling (40:46.960)
intelligent agent, what should the output of perception be? We have no idea. And how should
Leslie Kaelbling (40:52.560)
that hook up to the other stuff? We don't know. So I think the pressing question is,
Lex Fridman (40:59.040)
what kinds of structure can we build in that are like the moral equivalent of convolution
Leslie Kaelbling (41:03.520)
that will make a really awesome superstructure that then learning can kind of progress on
Leslie Kaelbling (41:09.440)
efficiently. I agree. Very compelling description of actually where we stand with the perception
Leslie Kaelbling (41:13.840)
problem. You're teaching a course on embodied intelligence. What do you think it takes to
Leslie Kaelbling (41:19.120)
build a robot with human level intelligence? I don't know if we knew we would do it.
Leslie Kaelbling (41:27.680)
If you were to, I mean, okay. So do you think a robot needs to have a self awareness,
Leslie Kaelbling (41:36.000)
consciousness, fear of mortality, or is it, is it simpler than that? Or is consciousness a simple
Leslie Kaelbling (41:44.160)
thing? Like, do you think about these notions? I don't think much about consciousness. Even
Leslie Kaelbling (41:50.880)
most philosophers who care about it will give you that you could have robots that are zombies,
Leslie Kaelbling (41:55.840)
right? That behave like humans, but are not conscious. And I, at this moment would be happy
Leslie Kaelbling (42:00.320)
enough with that. So I'm not really worried one way or the other. So the technical side,
Leslie Kaelbling (42:03.760)
you're not thinking of the use of self awareness. Well, but I, okay, but then what does self
Leslie Kaelbling (42:09.920)
awareness mean? I mean, that you need to have some part of the system that can observe other
Leslie Kaelbling (42:16.960)
parts of the system and tell whether they're working well or not. That seems critical.
Lex Fridman (42:21.200)
So does that count as, I mean, does that count as self awareness or not? Well, it depends on whether
Leslie Kaelbling (42:27.360)
you think that there's somebody at home who can articulate whether they're self aware. But clearly,
Leslie Kaelbling (42:33.120)
if I have like, you know, some piece of code that's counting how many times this procedure gets
Leslie Kaelbling (42:37.600)
executed, that's a kind of self awareness, right? So there's a big spectrum. It's clear you have to
Leslie Kaelbling (42:43.680)
have some of it. Right. You know, we're quite far away in many dimensions, but is there a direction
Leslie Kaelbling (42:48.160)
of research that's most compelling to you for, you know, trying to achieve human level intelligence
Leslie Kaelbling (42:54.720)
in our robots? Well, to me, I guess the thing that seems most compelling to me at the moment is this
Leslie Kaelbling (43:00.880)
question of what to build in and what to learn. Um, I think we're, we don't, we're missing a bunch
Leslie Kaelbling (43:10.160)
of ideas and, and we, you know, people, you know, don't you dare ask me how many years it's going to
Leslie Kaelbling (43:17.200)
be until that happens because I won't even participate in the conversation because I think
Leslie Kaelbling (43:22.320)
we're missing ideas and I don't know how long it's going to take to find them. So I won't ask you how
Leslie Kaelbling (43:26.400)
many years, but, uh, maybe I'll ask you what it, when you'll be sufficiently impressed that we've
Leslie Kaelbling (43:34.240)
achieved it. So what's, what's a good test of intelligence? Do you like the Turing test, the
Leslie Kaelbling (43:40.080)
natural language in the robotic space? Is there something where you would sit back and think,
Leslie Kaelbling (43:46.400)
Oh, that's, that's pretty impressive. Uh, as a test, as a benchmark, do you think about these
Leslie Kaelbling (43:52.000)
kinds of problems? No, I resist. I mean, I think all the time that we spend arguing about those
Leslie Kaelbling (43:57.760)
kinds of things could be better spent just making the robots work better. Uh, so you don't value
Leslie Kaelbling (44:03.520)
competition. So, I mean, there's a nature of benchmark benchmarks and datasets or Turing
Leslie Kaelbling (44:10.000)
test challenges where everybody kind of gets together and tries to build a better robot
Leslie Kaelbling (44:14.240)
cause they want to out compete each other. Like the DARPA challenge with the autonomous vehicles.
Lex Fridman (44:18.640)
Do you see the value of that or it can get in the way? I think it can get in the way. I mean,
Leslie Kaelbling (44:25.040)
some people, many people find it motivating. And so that's good. I find it anti motivating
Leslie Kaelbling (44:29.520)
personally. Uh, but I think what, I mean, I think you get an interesting cycle where for a contest,
Leslie Kaelbling (44:37.440)
a bunch of smart people get super motivated and they hack their brains out and much of what gets
Leslie Kaelbling (44:42.000)
done is just hacks, but sometimes really cool ideas emerge. And then that gives us something
Leslie Kaelbling (44:47.200)
to chew on after that. So I'm, it's not a thing for me, but I don't, I don't regret that other
Leslie Kaelbling (44:54.400)
people do it. Yeah. It's like you said with everything else that it makes us good. So jumping
Leslie Kaelbling (44:59.120)
topics a little bit, you started the journal of machine learning research and served as its editor
Leslie Kaelbling (45:05.440)
in chief. Uh, how did the publication come about and what do you think about the current publishing
Leslie Kaelbling (45:13.760)
model space in machine learning artificial intelligence? Okay, good. So it came about
Leslie Kaelbling (45:19.680)
because there was a journal called machine learning, which still exists, which was owned by
Leslie Kaelbling (45:24.000)
Cluer and there was, I was on the editorial board and we used to have these meetings annually where
Leslie Kaelbling (45:30.880)
we would complain to Cluer that it was too expensive for the libraries and that people
Leslie Kaelbling (45:34.640)
couldn't publish. And we would really like to have some kind of relief on those fronts and they would
Leslie Kaelbling (45:39.200)
always sympathize, but not do anything. So, uh, we just decided to make a new journal and, uh,
Leslie Kaelbling (45:46.960)
there was the journal of AI research, which has, was on the same model, which had been in existence
Leslie Kaelbling (45:52.720)
for maybe five years or so, and it was going on pretty well. So, uh, we just made a new journal.
Leslie Kaelbling (45:59.920)
It wasn't, I mean, um, I don't know, I guess it was work, but it wasn't that hard. So basically
Leslie Kaelbling (46:05.280)
the editorial board, probably 75% of the editorial board of, uh, machine learning resigned and we
Leslie Kaelbling (46:14.560)
founded the new journal, but it was sort of, it was more open. Yeah. Right. So it's completely
Leslie Kaelbling (46:21.760)
open. It's open access. Actually, uh, uh, I had a postdoc, George Conidaris who wanted to call
Leslie Kaelbling (46:28.960)
these journals free for all, uh, because there were, I mean, it both has no page charges and has
Leslie Kaelbling (46:36.240)
no, uh, uh, access restrictions. And the reason, and so lots of people, I mean, there were, there
Leslie Kaelbling (46:44.640)
were people who were mad about the existence of this journal who thought it was a fraud or
Leslie Kaelbling (46:48.960)
something. It would be impossible. They said to run a journal like this with basically, I mean,
Leslie Kaelbling (46:54.320)
for a long time, I didn't even have a bank account. Uh, I paid for the lawyer to incorporate and the
Leslie Kaelbling (47:00.320)
IP address and it just did cost a couple of hundred dollars a year to run. It's a little bit
Leslie Kaelbling (47:06.960)
more now, but not that much more, but that's because I think computer scientists are competent
Lex Fridman (47:13.920)
and autonomous in a way that many scientists and other fields aren't. I mean, at doing these kinds
Leslie Kaelbling (47:19.440)
of things, we already types out our own papers. We all have students and people who can hack a
Leslie Kaelbling (47:24.480)
website together in an afternoon. So the infrastructure for us was like, not a problem,
Lex Fridman (47:29.280)
but for other people in other fields, it's a harder thing to do. Yeah. And this kind of
Leslie Kaelbling (47:34.320)
open access journal is nevertheless one of the most prestigious journals. So it's not like, uh,
Leslie Kaelbling (47:41.600)
prestige and it can be achieved without any of the... Paper is not required for prestige.
Leslie Kaelbling (47:47.520)
Yeah. It turns out. Yeah. So on the review process side of actually a long time ago,
Leslie Kaelbling (47:53.600)
I don't remember when I reviewed a paper where you were also a reviewer. And I remember reading
Leslie Kaelbling (47:59.440)
your review being influenced by it and it was really well written. It influenced how I write
Leslie Kaelbling (48:04.080)
feature reviews. Uh, you disagreed with me actually. Uh, and you made it, uh, my review,
Lex Fridman (48:11.520)
but much better. So, but nevertheless, the review process, you know, has its, uh, flaws.
Lex Fridman (48:19.280)
And how do you think, what do you think works well? How can it be improved?
Lex Fridman (48:23.600)
So actually when I started JMLR, I wanted to do something completely different.
Lex Fridman (48:28.720)
And I didn't because it felt like we needed a traditional journal of record. And so we just
Leslie Kaelbling (48:34.800)
made JMLR be almost like a normal journal, except for the open access parts of it, basically. Um,
Leslie Kaelbling (48:43.200)
increasingly of course, publication is not even a sensible word. You can publish something by
Leslie Kaelbling (48:47.600)
putting it in an archive so I can publish everything tomorrow. So making stuff public
Leslie Kaelbling (48:54.240)
is, there's no barrier. We still need curation and evaluation. I don't have time to read all
Leslie Kaelbling (49:04.560)
of archive. And you could argue that kind of social thumbs upping of articles suffices,
Leslie Kaelbling (49:20.000)
right? You might say, Oh, heck with this. We don't need journals at all. We'll put everything
Leslie Kaelbling (49:24.880)
on archive and people will upvote and downvote the articles. And then your CV will say, Oh man,
Leslie Kaelbling (49:29.840)
he got a lot upvotes. So, uh, that's good. Um, but I think there's still
Leslie Kaelbling (49:39.040)
value in careful reading and commentary of things. And it's hard to tell when people are
Leslie Kaelbling (49:46.320)
upvoting and downvoting or arguing about your paper on Twitter and Reddit, whether they know
Lex Fridman (49:53.280)
what they're talking about, right? So then I have the second order problem of trying to decide whose
Leslie Kaelbling (49:57.760)
opinions I should value and such. So I don't know what I w if I had infinite time, which I don't,
Lex Fridman (50:04.480)
and I'm not going to do this because I really want to make robots work. But if I felt inclined to do
Leslie Kaelbling (50:10.160)
something more in the publication direction, I would do this other thing, which I thought about
Leslie Kaelbling (50:14.560)
doing the first time, which is to get together some set of people whose opinions I value and
Leslie Kaelbling (50:19.840)
who are pretty articulate. And I guess we would be public, although we could be private. I'm not sure.
Lex Fridman (50:25.200)
And we would review papers. We wouldn't publish them and you wouldn't submit them. We would just
Leslie Kaelbling (50:29.360)
find papers and we would write reviews and we would make those reviews public. And maybe if you,
Leslie Kaelbling (50:37.040)
you know, so we're Leslie's friends who review papers and maybe eventually if, if we, our opinion
Leslie Kaelbling (50:42.880)
was sufficiently valued, like the opinion of JMLR is valued, then you'd say on your CV that Leslie's
Leslie Kaelbling (50:48.560)
friends gave my paper a five star rating. And that would be just as good as saying, I got it,
Leslie Kaelbling (50:53.200)
so, you know, accepted into this journal. So I think, I think we should have good public commentary
Lex Fridman (51:01.840)
and organize it in some way, but I don't really know how to do it. It's interesting times.
Leslie Kaelbling (51:06.320)
The way you describe it actually is really interesting. I mean, we do it for movies,
Leslie Kaelbling (51:10.000)
imdb.com. There's experts, critics come in, they write reviews, but there's also
Leslie Kaelbling (51:16.000)
regular non critics, humans write reviews and they're separated.
Lex Fridman (51:19.840)
I like open review. The iClear process I think is interesting.
Leslie Kaelbling (51:29.280)
It's a step in the right direction, but it's still not as compelling as reviewing movies or
Leslie Kaelbling (51:35.760)
video games. I mean, it sometimes almost, it might be silly, at least from my perspective to say,
Lex Fridman (51:41.600)
but it boils down to the user interface, how fun and easy it is to actually perform the reviews,
Lex Fridman (51:46.720)
how efficient, how much you as a reviewer get street cred for being a good reviewer.
Leslie Kaelbling (51:54.400)
Those elements, those human elements come into play.
Leslie Kaelbling (51:57.200)
No, it's a big investment to do a good review of a paper and the flood of papers is out of control.
Leslie Kaelbling (52:04.000)
Right. So, you know, there aren't 3000 new, I don't know how many new movies are there in a year.
Leslie Kaelbling (52:08.480)
I don't know, but that's probably going to be less than how many machine learning papers are
Leslie Kaelbling (52:12.160)
in a year now. And I'm worried, you know, I, right. So I'm like an old person. So of course,
Leslie Kaelbling (52:21.760)
I'm going to say, things are moving too fast. I'm a stick in the mud. So I can say that,
Lex Fridman (52:28.720)
but my particular flavor of that is I think the horizon for researchers has gotten very short,
Leslie Kaelbling (52:35.760)
that students want to publish a lot of papers and there's a huge, there's value. It's exciting. And
Leslie Kaelbling (52:41.760)
there's value in that and you get patted on the head for it and so on. But, and some of that is
Leslie Kaelbling (52:49.200)
fine, but I'm worried that we're driving out people who would spend two years thinking about
Leslie Kaelbling (52:59.200)
something. Back in my day, when we worked on our thesis, we did not publish papers. You did your
Leslie Kaelbling (53:06.080)
thesis for years. You picked a hard problem and then you worked and chewed on it and did stuff
Lex Fridman (53:11.360)
and wasted time and for a long time. And when it was roughly, when it was done, you would write
Leslie Kaelbling (53:16.320)
papers. And so I don't know how to, and I don't think that everybody has to work in that mode,
Lex Fridman (53:23.520)
but I think there's some problems that are hard enough that it's important to have a long
Leslie Kaelbling (53:27.760)
research horizon. And I'm worried that we don't incentivize that at all at this point.
Leslie Kaelbling (53:33.040)
In this current structure. Yeah. So what do you see as, what are your hopes and fears about the
Leslie Kaelbling (53:41.840)
future of AI and continuing on this theme? So AI has gone through a few winters, ups and downs. Do
Lex Fridman (53:50.080)
you see another winter of AI coming? Are you more hopeful about making robots work, as you said?
Leslie Kaelbling (53:58.720)
I think the cycles are inevitable, but I think each time we get higher, right? I mean, so, you
Leslie Kaelbling (54:05.680)
know, it's like climbing some kind of landscape with a noisy optimizer. So it's clear that the,
Leslie Kaelbling (54:15.280)
you know, the deep learning stuff has made deep and important improvements. And so the high water
Leslie Kaelbling (54:22.560)
mark is now higher. There's no question. But of course, I think people are overselling and
Leslie Kaelbling (54:29.360)
eventually investors, I guess, and other people will look around and say, well, you're not quite
Leslie Kaelbling (54:37.040)
delivering on this grand claim and that wild hypothesis. It's probably, it's going to crash
Leslie Kaelbling (54:43.120)
some amount and then it's okay. I mean, but I don't, I can't imagine that there's like
Leslie Kaelbling (54:49.760)
some awesome monotonic improvement from here to human level AI. So in, you know, I have to ask
Leslie Kaelbling (54:58.320)
this question, I probably anticipate answers, the answers, but do you have a worry short term or
Leslie Kaelbling (55:05.520)
long term about the existential threats of AI and maybe short term, less existential, but more
Lex Fridman (55:15.280)
robots taking away jobs?
Leslie Kaelbling (55:17.120)
Well, actually, let me talk a little bit about utility. Actually, I had an interesting
Leslie Kaelbling (55:25.520)
conversation with some military ethicists who wanted to talk to me about autonomous weapons.
Lex Fridman (55:30.880)
And they're, they were interesting, smart, well educated guys who didn't know too much about AI
Leslie Kaelbling (55:37.440)
or machine learning. And the first question they asked me was, has your robot ever done
Leslie Kaelbling (55:41.440)
something you didn't expect? And I like burst out laughing because anybody who's ever done
Leslie Kaelbling (55:46.480)
something on the robot right knows that they don't do it. And what I realized was that their
Leslie Kaelbling (55:51.680)
model of how we program a robot was completely wrong. Their model of how we can program a robot
Leslie Kaelbling (55:56.960)
was like Lego mind storms, like, Oh, go forward a meter, turn left, take a picture, do this, do
Leslie Kaelbling (56:03.120)
that. And so if you have that model of programming, then it's true. It's kind of weird that your robot
Leslie Kaelbling (56:08.800)
would do something that you didn't anticipate. But the fact is, and actually, so now this is my
Leslie Kaelbling (56:13.760)
new educational mission. If I have to talk to non experts, I try to teach them the idea that
Leslie Kaelbling (56:20.640)
we don't operate, we operate at least one or maybe many levels of abstraction about that. And we say,
Leslie Kaelbling (56:26.400)
Oh, here's a hypothesis class, maybe it's a space of plans, or maybe it's a space of
Leslie Kaelbling (56:31.200)
classifiers or whatever. But there's some set of answers and an objective function. And then we
Leslie Kaelbling (56:36.400)
work on some optimization method that tries to optimize a solution solution in that class.
Lex Fridman (56:43.200)
And we don't know what solution is going to come out. Right. So I think it's important to
Leslie Kaelbling (56:47.600)
communicate that. So I mean, of course, probably people who listen to this, they, they know that
Leslie Kaelbling (56:52.160)
lesson. But I think it's really critical to communicate that lesson. And then lots of people
Leslie Kaelbling (56:56.320)
are now talking about, you know, the value alignment problem. So you want to be sure as
Leslie Kaelbling (57:01.840)
robots or software systems get more competent, that their objectives are aligned with your
Leslie Kaelbling (57:07.840)
objectives, or that our objectives are compatible in some way, or we have a good way of mediating
Leslie Kaelbling (57:13.760)
when they have different objectives. And so I think it is important to start thinking in terms
Leslie Kaelbling (57:19.120)
like, you don't have to be freaked out by the robot apocalypse, to accept that it's important
Leslie Kaelbling (57:25.360)
to think about objective functions of value alignment. Yes. And that you have to really
Leslie Kaelbling (57:30.080)
everyone who's done optimization knows that you have to be careful what you wish for that,
Leslie Kaelbling (57:34.160)
you know, sometimes you get the optimal solution, and you realize, man, that was that objective was
Leslie Kaelbling (57:38.800)
wrong. So pragmatically, in the shortest term, it seems to me that that that those are really
Leslie Kaelbling (57:47.040)
interesting and critical questions. And the idea that we're going to go from being people who
Leslie Kaelbling (57:50.960)
engineer algorithms to being people who engineer objective functions. I think that's, that's
Leslie Kaelbling (57:56.480)
definitely going to happen. And that's going to change our thinking and methodology. And so we're
Leslie Kaelbling (58:00.560)
gonna you started at Stanford philosophy, that's where she could be. And I will go back to
Leslie Kaelbling (58:05.920)
philosophy maybe. Well, I mean, they're mixed together, because because, as we also know,
Leslie Kaelbling (58:11.920)
as machine learning people, right? When you design, in fact, this is the lecture I gave in
Leslie Kaelbling (58:16.160)
class today, when you design an objective function, you have to wear both hats, there's
Leslie Kaelbling (58:21.600)
the hat that says, what do I want? And there's the hat that says, but I know what my optimizer
Leslie Kaelbling (58:26.080)
can do to some degree. And I have to take that into account. So it's it's always a trade off,
Lex Fridman (58:31.840)
and we have to kind of be mindful of that. The part about taking people's jobs, I understand
Leslie Kaelbling (58:38.560)
that that's important. I don't understand sociology or economics or people very well. So I
Leslie Kaelbling (58:45.440)
don't know how to think about that. So that's Yeah, so there might be a sociological aspect
Leslie Kaelbling (58:50.160)
there, the economic aspect that's very difficult to think about. Okay. I mean, I think other people
Leslie Kaelbling (58:54.640)
should be thinking about it. But I'm just that's not my strength. So what do you think is the most
Lex Fridman (58:58.480)
exciting area of research in the short term, for the community and for your for yourself?
Leslie Kaelbling (59:03.600)
Well, so I mean, there's the story I've been telling about how to engineer intelligent robots.
Lex Fridman (59:10.160)
So that's what we want to do. We all kind of want to do well, I mean, some set of us want to do this.
Lex Fridman (59:16.160)
And the question is, what's the most effective strategy? And we've tried it. And there's a bunch
Leslie Kaelbling (59:20.560)
of different things you could do at the extremes, right? One super extreme is, what's the most
Leslie Kaelbling (59:25.360)
effective strategy? And there's a bunch of different things you could do at the extremes,
Leslie Kaelbling (59:29.600)
right? One super extreme is, we do introspection, and we write a program. Okay, that has not worked
Leslie Kaelbling (59:35.920)
out very well. Another extreme is we take a giant bunch of neural goo, and we try and train it up to
Leslie Kaelbling (59:41.280)
do something. I don't think that's going to work either. So the question is, what's the middle
Leslie Kaelbling (59:46.960)
ground? And, and again, this isn't a theological question or anything like that. It's just like,
Leslie Kaelbling (59:54.240)
what's the middle ground? And I think it's clear, it's a combination of learning, to me, it's clear,
🔗 相关节目