Leslie Kaelbling: Reinforcement Learning, Planning, and Robotics
嘉宾：Leslie Kaelbling
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
§01Transcript · 00:00–19:06
The following is a conversation with Leslie Kaelbling. She's a roboticist and professor at
MIT. She's recognized for her work in reinforcement learning, planning, robot navigation, and several
other topics in AI. She won the IJCAI Computers and Thought Award and was the editor-in-chief
of the prestigious Journal of Machine Learning Research. This conversation is part of the
Artificial Intelligence Podcast at MIT and beyond. If you enjoy it, subscribe on YouTube, iTunes,
or simply connect with me on Twitter @LexFriedman, spelled F-R-I-D. And now,
here's my conversation with Leslie Kaelbling.
- What made me get excited about AI, I can say that, is I read Gödel, Escher, Bach when I was
in high school. That was pretty formative for me because it exposed the interestingness of
primitives and combination and how you can make complex things out of simple parts,
and ideas of AI and what kinds of programs might generate intelligent behavior.
- So you first fell in love with AI reasoning logic versus robots?
- Yeah, the robots came because my first job, so I finished an undergraduate degree in philosophy
at Stanford and was about to finish a master's in computer science, and I got hired at SRI
in their AI lab. And they were building a robot. It was a kind of a follow-on to Shakey,
but all the Shakey people were not there anymore. And so my job was to try to get this robot to do
stuff, and that's really kind of what got me interested in robots.
- So maybe taking a small step back, your bachelor's in Stanford in philosophy,
did a master's and PhD in computer science, but the bachelor's in philosophy. So what was that
journey like? What elements of philosophy do you think you bring to your work in computer science?
- So it's surprisingly relevant. So part of the reason that I didn't do a computer science
undergraduate degree was that there wasn't one at Stanford at the time, but that there's part
of philosophy, and in fact, Stanford has a special sub-major in something called now
symbolic systems, which is logic, model, theory, formal semantics of natural language. And so that's
actually a perfect preparation for work in AI and computer science.
- That's kind of interesting. So if you were interested in artificial intelligence,
what kind of majors were people even thinking about taking? What is it in neuroscience?
So besides philosophies, what were you supposed to do if you were
fascinated by the idea of creating intelligence?
- There weren't enough people who did that for that even to be a conversation. I mean,
I think probably philosophy. I mean, it's interesting in my class, my graduating class of
undergraduate philosophers, probably maybe slightly less than half went on in computer science,
slightly less than half went on in law, and like one or two went on in philosophy.
So it was a common kind of connection.
- Do you think AI researchers have a role, be part-time philosophers, or should they stick
to the solid science and engineering without sort of taking the philosophizing tangents? I mean,
you work with robots, you think about what it takes to create intelligent beings. Aren't you
the perfect person to think about the big picture philosophy at all?
- The parts of philosophy that are closest to AI, I think, or at least the closest to AI that I think
about are stuff like belief and knowledge and denotation and that kind of stuff. And that's,
you know, it's quite formal and it's like just one step away from
the kinds of computer science work that we do kind of routinely.
I think that there are important questions still about what you can do with a machine and what you
can't and so on. Although at least my personal view is that I'm completely a materialist and
I don't think that there's any reason why we can't make a robot be behaviorally indistinguishable
from a human. And the question of whether it's distinguishable internally, whether it's a zombie
or not in philosophy terms, I actually don't, I don't know, and I don't know if I care too much
about that. - Right, but there is philosophical notions, they're mathematical and philosophical
because we don't know so much, of how difficult that is, how difficult is the perception problem,
how difficult is the planning problem, how difficult is it to operate in this world successfully.
Because our robots are not currently as successful as human beings in many tasks,
the question about the gap between current robots and human beings borders a little bit on
philosophy. You know, the expanse of knowledge that's required to operate in this world,
and the ability to form common sense knowledge, the ability to reason about uncertainty,
much of the work you've been doing, there's open questions there that, I don't know,
require to activate a certain big picture view. - To me that doesn't seem like a philosophical
gap at all. To me, there is a big technical gap, there's a huge technical gap, but I don't see any
reason why it's more than a technical gap. - Perfect. So, when you mentioned AI, you mentioned
SRI, and maybe, can you describe to me when you first fell in love with robotics, the robots,
were inspired, which, so you mentioned Flaky or Shaky, Shaky Flaky, and what was the robot that
first captured your imagination, what's possible? - Right, well, so the first robot I worked with was
Flaky. Shaky was a robot that the SRI people had built, but by the time, I think when I arrived,
it was sitting in a corner of somebody's office dripping hydraulic fluid into a pan.
But it's iconic, and really, everybody should read the Shaky Tech Report, because it has so
many good ideas in it. I mean, they invented A*Search and symbolic planning and learning
macro operators. They had low-level kind of configuration space planning for their robot,
they had vision, they had all, this, the basic ideas of a ton of things. - Can you take,
step by, Shaky have arms, what was the job, what was the goal? - Shaky was a mobile robot,
but it could push objects, and so it would move things around. - With which actuator, with arms?
- With itself, with its base. - Okay, great. - So it could, but it, and they had painted the base
boards black, so it used vision to localize itself in a map, it detected objects, it could detect
objects that were surprising to it, it would plan and replan based on what it saw, it reasoned about
whether to look and take pictures. I mean, it really had the basics of so many of the things
that we think about now. - How did it represent the space around it? - So it had representations
at a bunch of different levels of abstraction, so it had, I think, a kind of an occupancy grid
of some sort at the lowest level. At the high level, it was abstract, symbolic kind of rooms
and connectivity. - So where does Flaky come in? - Yeah, okay, so I showed up at SRI, and the,
we were building a brand new robot. As I said, none of the people from the previous project were
kind of there or involved anymore, so we were kind of starting from scratch, and my advisor
was Stan Rosenstein, he ended up being my thesis advisor, and he was motivated by this idea of
situated computation or situated automata, and the idea was that the tools of logical reasoning were
important, but possibly only for the engineers or designers to use in the analysis of a system,
but not necessarily to be manipulated in the head of the system itself, right? So I might use logic
to prove a theorem about the behavior of my robot, even if the robot's not using logic in its head
to prove theorems, right? So that was kind of the distinction. And so the idea was to kind of use
those principles to make a robot do stuff, but a lot of the basic things we had to kind of
learn for ourselves, 'cause I had zero background in robotics, I didn't know anything about control,
I didn't know anything about sensors, so we reinvented a lot of wheels on the way to getting
that robot to do stuff. - Do you think that was an advantage
or a hindrance? - Oh, no, I mean, I'm big in favor of
wheel reinvention, actually. I mean, I think you learn a lot by doing it. It's important, though,
to eventually have the pointers so that you can see what's really going on, but I think you can
appreciate much better the good solutions once you've messed around a little bit on your own
and found a bad one. - Yeah, I think you mentioned
reinventing reinforcement learning and referring to rewards as pleasures, a pleasure, I think,
which I think is a nice name for it. - Yeah, it seemed good to me.
- It's more fun, almost. Do you think you could tell the history of AI, machine learning,
reinforcement learning, how you think about it from the '50s to now?
- One thing is that it oscillates, right? So things become fashionable and then they go out
and then something else becomes cool and then it goes out and so on. So there's some interesting
sociological process that actually drives a lot of what's going on. Early days was kind of
cybernetics and control, right? And the idea of homeostasis, right? People who made these robots
that could, I don't know, try to plug into the wall when they needed power and then come loose
and roll around and do stuff. And then I think over time, the thought, well, that was inspiring,
but people said, no, no, no, we want to get maybe closer to what feels like real intelligence or
human intelligence. And then maybe the expert systems people tried to do that, but maybe a
little too superficially, right? So, oh, we get the surface understanding of what intelligence is
like because I understand how a steel mill works and I can try to explain it to you and you can
write it down in logic and then we can make a computer infer that. And then that didn't work
out. But what's interesting, I think, is when a thing starts to not be working very well,
it's not only do we change methods, we change problems, right? So it's not like we have better
ways of doing the problem that the expert systems people were trying to do. We have no ways of
trying to do that problem. Oh yeah, I know, I think, or maybe a few, but we kind of give up
on that problem and we switch to a different problem and we work that for a while and we make
progress. - As a broad community.
- As a community, yeah. - And there's a lot of people
who would argue you don't give up on the problem, it's just you decrease the number of people working
on it. You almost kind of like put it on the shelf, say, we'll come back to this 20 years later.
- Yeah, I think that's right. Or you might decide that it's malformed.
- Like you might say, it's wrong to just try to make something that does superficial symbolic
reasoning behave like a doctor. You can't do that until you've had the sensory motor experience
of being a doctor or something, right? So there's arguments that say that that problem was not well
formed, or it could be that it is well formed, but we just weren't approaching it well.
- So you mentioned that your favorite part of logic and symbolic systems is that they give
short names for large sets. So there is some use to this, the use to symbolic reasoning. So
looking at expert systems and symbolic computing, what do you think are the roadblocks that were
hit in the '80s and '90s? - Ah, okay. So right, so the fact that I'm
not a fan of expert systems doesn't mean that I'm not a fan of some kinds of symbolic reasoning.
Let's see, roadblocks. Well, the main roadblock, I think, was that the idea that humans could
articulate their knowledge effectively into some kind of logical statements.
- So it's not just the cost, the effort, but really just the capability of doing it.
- Right, because we're all experts in vision, right? But totally don't have introspective
access into how we do that, right? And it's true that, I mean, I think the idea was, well, of
course, even people then would know, of course, I wouldn't ask you to please write down the rules
that you use for recognizing a water bottle. That's crazy. And everyone understood that. But
we might ask you to please write down the rules you use for deciding, I don't know, what tie to
put on or how to set up a microphone or something like that. But even those things, I think people
maybe, I think what they found, I'm not sure about this, but I think what they found was that the
so-called experts could give explanations that sort of post hoc explanations for how and why
they did things, but they weren't necessarily very good. And then they depended on maybe some
kinds of perceptual things, which again, they couldn't really define very well. So I think
fundamentally, I think that the underlying problem with that was the assumption that people could
articulate how and why they make their decisions. - Right, so it's almost encoding the knowledge,
converting from expert to something that a machine could understand and reason with.
- No, no, no, not even just encoding, but getting it out of you.
- Just... - Right? Not writing it,
I mean, yes, hard also to write it down for the computer, but I don't think that people can
produce it. You can tell me a story about why you do stuff, but I'm not so sure that's the why.
- Great. So there are still on the hierarchical planning side,
places where symbolic reasoning is very useful. So as you've talked about, so...
- Right, so don't... - Where's the gap?
- Yeah, okay, good. So saying that humans can't provide a description of their reasoning processes,
that's okay, fine, but that doesn't mean that it's not good to do reasoning of various styles
inside a computer. Those are just two orthogonal points. So then the question is, what kind of
reasoning should you do inside a computer? And the answer is, I think you need to do all different
kinds of reasoning inside a computer, depending on what kinds of problems you face.
- I guess the question is, what kind of things can you encode symbolically so you can reason about?
- I think the idea about, and even symbolic, I don't even like that terminology,
'cause I don't know what it means technically and formally. I do believe in abstractions.
So abstractions are critical, right? You cannot reason at completely fine grain about everything
in your life, right? You can't make a plan at the level of images and torques for getting a PhD.
So you have to reduce the size of the state space and you have to reduce the horizon
if you're gonna reason about getting a PhD or even buying the ingredients to make dinner.
And so how can you reduce the spaces and the horizon of the reasoning you have to do? And
the answer is abstraction, spatial abstraction, temporal abstraction. I think abstraction along
the lines of goals is also interesting. Like you might, or well, abstraction and decomposition.
Goals is maybe more of a decomposition thing. So I think that's where these kinds of, if you want
to call it symbolic or discrete models come in. You talk about a room of your house instead of
your pose. You talk about doing something during the afternoon instead of at 2.54.
And you do that because it makes your reasoning problem easier. And also because
you don't have enough information to reason in high fidelity about your pose of your elbow at
2.35 this afternoon anyway. - Right. When you're trying to get a PhD.
- Right. Or when you're doing anything really. - Yeah, okay.
- Except for at that moment. At that moment, you do have to reason about the pose of your elbow,
maybe. But then maybe you do that in some continuous joint space kind of model. So again,
my biggest point about all of this is that there should be, that dogma is not the thing, right?
It shouldn't be that I am in favor against symbolic reasoning and you're in favor against
neural networks. It should be that just computer science tells us what the right answer to all
these questions is if we were smart enough to figure it out. - Well, yeah. When you try to
actually solve the problem with computers, the right answer comes out. You mentioned abstractions.
I mean, neural networks form abstractions or rather there's automated ways to form abstractions.
- Absolutely. - And there's
expert driven ways to form abstractions and expert human driven ways. And humans just seems to be
way better at forming abstractions currently and certain problems. So when you're referring to
2.45 p.m. versus afternoon, how do we construct that taxonomy? Is there any room for automated
construction of such abstractions? - Oh, I think eventually, yeah. I mean,
I think when we get to be better machine learning engineers, we'll build algorithms that
build awesome abstractions. - That are useful in this kind
of way that you're describing. - Yeah.
- Yeah. So let's then step from the abstraction discussion and let's talk about BOM MDPs,
partially observable Markov decision processes. So uncertainty. So first,
what are Markov decision processes? - What are Markov decision processes?
- And maybe how much of our world can be models and MDPs? How much, when you wake up in the
morning and you're making breakfast, do you think of yourself as an MDP? So how do you think about
MDPs and how they relate to our world? - Well, so there's a stance question, right?
So a stance is a position that I take with respect to a problem. So I, as a researcher or a person
who designs systems, can decide to make a model of the world around me in some terms. So I take
this messy world and I say, I'm going to treat it as if it were a problem of this formal kind,
and then I can apply solution concepts or algorithms or whatever to solve that formal
thing, right? So of course the world is not anything. It's not an MDP or a POMDP. I don't
§02Transcript · 19:11–38:40
know what it is, but I can model aspects of it in some way or some other way. And when I model some
aspect of it in a certain way, that gives me some set of algorithms I can use. - You can model the
world in all kinds of ways. Some are more accepting of uncertainty, more easily modeling uncertainty
of the world. Some really force the world to be deterministic. And so certainly MDPs model the
uncertainty of the world. - Yes. Model some uncertainty. They model not present state
uncertainty, but they model uncertainty in the way the future will unfold. - Right. So what are
Markov decision processes? - Okay, so Markov decision process is a model. It's a kind of a
model that you could make that says, I know completely the current state of my system.
And what it means to be a state is that I have all the information right now that will let me make
predictions about the future as well as I can. So that remembering anything about my history
wouldn't make my predictions any better. But then it also says that then I can take some actions
that might change the state of the world and that I don't have a deterministic model of those
changes. I have a probabilistic model of how the world might change. It's a useful model for some
kinds of systems. I think it's a, I mean, it's certainly not a good model for most problems.
I think because for most problems, you don't actually know the state. For most problems,
it's partially observed. So that's now a different problem class. - So, okay, that's where the
POMDPs, the partially observable Markov decision processes step in. So how do they address the
fact that you can't observe most, you have incomplete information about most of the world
around you? - Right. So now the idea is we still kind of postulate that there exists a state. We
think that there is some information about the world out there such that if we knew that we
could make good predictions, but we don't know the state. And so then we have to think about how,
but we do get observations. Maybe I get images or I hear things or I feel things, and those might be
local or noisy. And so therefore they don't tell me everything about what's going on. And then I
have to reason about, given the history of actions I've taken and observations I've gotten,
what do I think is going on in the world? And then given my own kind of uncertainty about what's
going on in the world, I can decide what actions to take. - And so difficult is this problem of
planning under uncertainty in your view, in your long experience of modeling the world,
trying to deal with this uncertainty in, especially in real world systems. - Optimal planning for even
discrete POMDPs can be undecidable depending on how you set it up. And so lots of people say,
I don't use POMDPs because they are intractable. And I think that that's a kind of a very funny
thing to say, because the problem you have to solve is the problem you have to solve.
So if the problem you have to solve is intractable, that's what makes us AI people,
right? So we solve, we understand that the problem we're solving is wildly intractable,
that we can't, we will never be able to solve it optimally. At least I don't. Yeah, right. So
later we can come back to an idea about bounded optimality and something. But anyway,
we can't come up with optimal solutions to these problems. So we have to make approximations,
approximations in modeling, approximations in solution algorithms and so on. And so
I don't have a problem with saying, yeah, my problem actually, it is POMDP in continuous
space with continuous observations and it's so computationally complex, I can't even think about
it's, you know, big O, whatever. But that doesn't prevent me from, it helps me,
gives me some clarity to think about it that way. And to then take steps to make approximation after
approximation to get down to something that's like computable in some reasonable time.
When you think about optimality, you know, the community broadly has shifted on that,
I think a little bit in how much they value the idea of optimality, of chasing an optimal solution.
How has your views of chasing an optimal solution changed over the years when you work with robots?
That's interesting. I think we have a little bit of a methodological crisis, actually,
from the theoretical side. I mean, I do think that theory is important and that right now we're not
doing much of it. So there's lots of empirical hacking around and training this and doing that
and reporting numbers, but is it good? Is it bad? We don't know. It's very hard to say things.
And if you look at like computer science theory, so people talked for a while, everyone was about
solving problems optimally or completely. And then there were interesting relaxations, right? So
people look at, oh, can I, are there regret bounds or can I do some kind of, you know,
approximation? Can I prove something that I can approximately solve this problem or that I get
closer to the solution as I spend more time and so on? What's interesting, I think, is that we don't
have good approximate solution concepts for very difficult problems, right? I like to, you know,
I like to say that I'm interested in doing a very bad job of very big problems.
- That's a good quote.
- Right. So very bad job of very big problems. I like to do that. But I wish I could say
something. I wish I had a, I don't know, some kind of a formal solution concept that I could use to
say, oh, this algorithm actually, it gives me something. Like I know what I'm going to get.
I can do something other than just run it and get out 6.7.
- That notion is still somewhere deeply compelling to you. The notion that you can say,
you can drop thing on the table says this, you can expect this algorithm will give me some good
results.
- I hope there's, I hope science will, I mean, there's engineering and there's science. I think
that they're not exactly the same. And I think right now we're making huge engineering, like,
leaps and bounds. So the engineering is running away ahead of the science, which is cool and
often how it goes, right? So we're making things and nobody knows how and why they work, roughly.
But we need to turn that into science.
- There's some form, it's, yeah, there's some room for formalizing.
- We need to know what the principles are. Why does this work? Why does that not work? I mean,
for a while people built bridges by trying, but now we can often predict whether it's going to
work or not without building it. Can we do that for learning systems or for robots?
- So your hope is from a materialistic perspective that
intelligence, artificial intelligence systems, robots, are just fancier bridges.
Belief space. What's the difference between belief space and state space? So you mentioned MDPs,
POMDPs, reasoning about, you sense the world, there's a state. What's this belief space idea?
- Yeah, that sounds so good.
- That sounds good. So belief space, that is, instead of thinking about what's the state of
the world and trying to control that as a robot, I think about what is the space of beliefs that I
could have about the world? What's, if I think of a belief as a probability distribution of ways the
world could be, a belief state as a distribution, and then my control problem, if I'm reasoning
about how to move through a world I'm uncertain about, my control problem is actually the problem
of controlling my beliefs. So I think about taking actions, not just what effect they'll
have on the world outside, but what effect they'll have on my own understanding of the world outside.
And so that might compel me to ask a question or look somewhere to gather information,
which may not really change the world state, but it changes my own belief about the world.
- That's a powerful way to empower the agent to reason about the world, to explore the world.
What kind of problems does it allow you to solve to consider belief space versus just state space?
- Well, any problem that requires deliberate information gathering, right? So if,
in some problems, like chess, there's no uncertainty, or maybe there's uncertainty
about the opponent, there's no uncertainty about the state. And some problems there's
uncertainty, but you gather information as you go, right? You might say, "Oh, I'm driving my
autonomous car down the road, and it doesn't know perfectly where it is, but the light hours are all
going all the time, so I don't have to think about whether to gather information." But if you're a
human driving down the road, you sometimes look over your shoulder to see what's going on behind
you in the lane, and you have to decide whether you should do that now. And you have to trade off
the fact that you're not seeing in front of you, and you're looking behind you, and how valuable
is that information, and so on. And so to make choices about information gathering,
you have to reason in belief space. Also, I mean, also to just take into account your own
uncertainty before trying to do things. So you might say, "If I understand where I'm standing
relative to the door jam pretty accurately, then it's okay for me to go through the door. But if
I'm really not sure where the door is, then it might be better to not do that right now."
The degree of your uncertainty about the world is actually part of the thing you're trying to
optimize in forming the plan, right? That's right.
So this idea of a long horizon of planning for a PhD, or just even how to get out of the house,
or how to make breakfast. You show this presentation of the WTF, where's the fork,
of robot looking at a sink. And can you describe how we plan in this world,
of this idea of hierarchical planning we've mentioned? So yeah, how can a robot hope to
plan about something with such a long horizon, where the goal is quite far away?
People, since probably reasoning began, have thought about hierarchical reasoning.
The temporal hierarchy in particular. Well, there's spatial hierarchy, but let's talk
about temporal hierarchy. So you might say, "Oh, I have this long execution I have to do,
but I can divide it into some segments abstractly." So maybe you have to get out of the house,
I have to get in the car, I have to drive, and so on. And so you can plan. If you can build
abstractions, so this, we started out by talking about abstractions, and we're back to that now.
If you can build abstractions in your state space, and abstractions, sort of temporal abstractions,
then you can make plans at a high level. And you can say, "I'm going to go to town, and then I'll
have to get gas, and then I can go here, and I can do this other thing." And you can reason about
the dependencies and constraints among these actions, again, without thinking about the complete
details. What we do in our hierarchical planning work is then say, "All right, I make a plan at a
high level of abstraction. I have to have some reason to think that it's feasible without
working it out in complete detail." And that's actually the interesting step. I always like to
talk about walking through an airport. Like, you can plan to go to New York and arrive at the
airport, and then find yourself an office building later. You can't even tell me in advance what your
plan is for walking through the airport. Partly because you're too lazy to think about it, maybe,
but partly also because you just don't have the information. You don't know what gate you're
landing in, or what people are going to be in front of you, or anything. So there's no point
in planning in detail. But you have to have -- you have to make a leap of faith that you can figure
it out once you get there. And it's really interesting to me how you arrive at that.
How do you -- so you have learned over your lifetime to be able to make some kinds of
predictions about how hard it is to achieve some kinds of sub-goals. And that's critical. Like,
you would never plan to fly somewhere if you couldn't -- didn't have a model of how hard it
was to do some of the intermediate steps. So one of the things we're thinking about now is,
how do you do this kind of very aggressive generalization to situations that you haven't
been in and so on, to predict how long will it take to walk through the Kuala Lumpur airport?
Like, you could give me an estimate and it wouldn't be crazy. And you have to have an estimate of that
in order to make plans that involve walking through the Kuala Lumpur airport, even if you
don't need to know it in detail. So I'm really interested in these kinds of abstract models and
how do we acquire them. But once we have them, we can use them to do hierarchical reasoning,
which I think is very important. Yeah, there's this notion of goal regression and
pre-image backchaining, this idea of starting at the goal and just forming these big clouds
of states. I mean, it's almost like saying to the airport, you know, you know, once you show up
to the airport, that you're like a few steps away from the goal. So like, thinking of it this way,
it's kind of interesting. I don't know if you have sort of further comments on that,
of starting at the goal. Yeah, I mean, it's interesting that Simon, Herb Simon, back in the
early days of AI, talked a lot about means-ends reasoning and reasoning back from the goal.
There's a kind of an intuition that people have that the number of, that state space is big,
the number of actions you could take is really big. So if you say, here I sit and I want to
search forward from where I am, what are all the things I could do? That's just overwhelming.
If you say, if you can reason at this other level and say, here's what I'm hoping to achieve,
what could I do to make that true? That somehow the branching is smaller. Now,
what's interesting is that like in the AI planning community, that hasn't worked out. In the class of
problems that they look at and the methods that they tend to use, it hasn't turned out that it's
better to go backward. It's still kind of my intuition that it is, but I can't prove that
to you right now. Right. I share your intuition, at least for us mere humans. Speaking of which,
when you maybe now we take it and take a little step into that philosophy circle,
how hard would it, when you think about human life, you give those examples often,
how hard do you think it is to formulate human life as a planning problem or aspects of
human life? So when you look at robots, you're often trying to think about object manipulation,
tasks, about moving a thing. When you take a slight step outside the room, let the robot
leave and go get lunch, or maybe try to pursue more fuzzy goals. How hard do you think is that
problem? If you were to try to maybe put another way, try to formulate human life as a planning
problem. Well, that would be a mistake. I mean, it's not all a planning problem, right? I think
it's really, really important that we understand that you have to put together pieces and parts
that have different styles of reasoning and representation and learning. I think it seems
probably clear to anybody that it can't all be this or all be that. Brains aren't all like this
or all like that, right? They have different pieces and parts and substructure and so on.
So I don't think that there's any good reason to think that there's going to be like one
true algorithmic thing that's going to do the whole job.
Just a bunch of pieces together designed to solve a bunch of specific problems.
Or maybe styles of problems. I mean, there's probably some reasoning that needs to go on
in image space. I think, again, there's this model-based versus model-free idea, right? So
in reinforcement learning, people talk about, "Oh, should I learn? I could learn a policy,
just straight up a way of behaving. I could learn it's popular, learn a value function,
that's some kind of weird intermediate ground. Or I could learn a transition model, which tells me
something about the dynamics of the world." If I take a tra- imagine that I learn a transition
model and I couple it with a planner and I draw a box around that, I have a policy again. It's just
stored a different way. But it's just as much of a policy as the other policy. It's just I've made,
I think, the way I see it is it's a time-space trade-off in computation. Right? A more overt
policy representation. Maybe it takes more space, but maybe I can compute quickly what action I
should take. On the other hand, maybe a very compact model of the world dynamics plus a
planner lets me compute what action to take too, just more slowly. There's no, I don't, I mean,
I don't think there's no argument to be had. It's just like a question of what form of computation
is best for us. - For the various sub-problems. - Right. So, and so like learning to do algebra
manipulations for some reason is, I mean, that's probably gonna want naturally a sort of a
different representation than writing a unicycle. At the time constraints on the unicycle are
serious. The space is maybe smaller. I don't know. But so I. - And there could be the more human
sides of falling in love, having a relationship that might be another. - Yeah, I have no idea.
- How to model that. Yeah. Let's first solve the algebra and the object manipulation.
What do you think is harder, perception or planning? - Perception. That's why. - Understanding.
That's why. So what do you think is so hard about perception,
about understanding the world around you? - Well, I mean, I think the big question
is representational. Hugely the question is representation. So perception has made
great strides lately, right? And we can classify images and we can
play certain kinds of games and predict how to steer the car and all this sort of stuff.
I don't think we have a very good idea of what perception should deliver, right? So if you,
if you believe in modularity, okay, there's a very strong view which says
we shouldn't build in any modularity. We should make a giant, gigantic neural network,
train it end to end to do the thing. And that's the best way forward. And it's hard to argue
§03Transcript · 38:47–58:51
with that except on a sample complexity basis, right? So you might say, oh, well, if I want to
do end to end reinforcement learning on this giant, giant neural network, it's going to take
a lot of data and a lot of like broken robots and stuff. So then the only answer is to say, okay,
we have to build something in, build in some structure or some bias. We know from theory of
machine learning, the only way to cut down the sample complexity is to kind of cut down,
somehow cut down the hypothesis space. You can do that by building in bias. There's all kinds
of reasons to think that nature built bias into humans. Convolution is a bias. It's a very strong
bias and it's a very critical bias. So my own view is that we should look for more things that are
like convolution, but that address other aspects of reasoning, right? So convolution helps us a
lot with a certain kind of spatial reasoning. That's quite close to the imaging. I think
there's other ideas like that. Maybe some amount of forward search, maybe some notions of abstraction,
maybe the notion that objects exist. Actually, I think that's pretty important. And a lot of people
won't give you that to start with. Right? - So almost like a convolution in the,
in the object, semantic object space or some kind of, some kind of ideas in there.
- That's right. And people are starting, like the graph, graph convolutions are an idea that
are related to relational representations. And so, so I think there are, so you, I've come far
afield from perception, but I think, I think the thing that's going to make perception that kind
of the next step is actually understanding better what it should produce. Right? So what are we
going to do with the output of it? Right? It's fine when what we're going to do with the output
is severe. It's less clear when we're just trying to make a one integrated, intelligent agent. What
should the output of perception be? We have no idea. And how should that hook up to the other
stuff? We don't know. So I think the pressing question is what kinds of structure can we build
in that are like the moral equivalent of convolution that will make a really awesome
superstructure that then learning can kind of progress on efficiently.
- I agree. Very compelling description of actually where we stand with the perception problem.
You're teaching a course on embodying intelligence. What do you think it takes to
build a robot with human level intelligence? - I don't know if we knew we would do it.
- If you were to, I mean, okay. So do you think a robot needs to have a self-awareness,
a consciousness, fear of mortality, or is it simpler than that? Or is consciousness a simple
thing? Do you think about these notions? - I don't think much about consciousness.
Even most philosophers who care about it will give you that you could have robots that are zombies,
right? That behave like humans, but are not conscious. And I, at this moment,
would be happy enough with that. So I'm not really worried one way or the other.
- So on the technical side, you're not thinking of the use of self-awareness?
- Well, but I, okay. But then what does self-awareness mean? I mean,
that you need to have some part of the system that can observe other parts of the system and
tell whether they're working well or not. That seems critical. So does that count as, I mean,
does that count as self-awareness or not? Well, it depends on whether you think that there's
somebody at home who can articulate whether they're self-aware. But clearly, if I have
some piece of code that's counting how many times this procedure gets executed,
that's a kind of self-awareness, right? So there's a big spectrum. It's clear you have to have some
of it. - Right. We're quite far away,
how many dimensions, but is there a direction of research that's most compelling to you for
trying to achieve human-level intelligence in our robots?
- Well, to me, I guess the thing that seems most compelling to me at the moment is this
question of what to build in and what to learn. I think we're missing a bunch of ideas and
we, you know, people, you know, don't you dare ask me how many years it's going to be until that
happens because I won't even participate in the conversation because I think 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 many years,
but maybe I'll ask you when you'll be sufficiently impressed that we've achieved it. So what's a good
test of intelligence? Do you like the Turing test, the natural language, and the robotic space? Is
there something where you would sit back and think, "Oh, that's pretty impressive as a test,
as a benchmark." Do you think about these kinds of problems? - No, I resist. I mean, I think all
the time that we spend arguing about those kinds of things could be better spent just making the
robots work better. - You don't value competition. So, I mean, there's a nature of benchmarks and
data sets or Turing test challenges where everybody kind of gets together and tries to build a better
robot because they want to out-compete each other. Like the DARPA challenge with the autonomous
vehicles. Do you see the value of that or it can get in the way? - I think it can get in the way.
Many people find it motivating and so that's good. I find it anti-motivating personally.
But I think you get an interesting cycle where for a contest, a bunch of smart people get super
motivated and they hack their brains out. And much of what gets done is just hacks, but sometimes
really cool ideas emerge. And then that gives us something to chew on after that. So, it's not a
thing for me, but I don't regret that other people do it. - Yeah, it's like you said, with everything
else, the mix is good. So, jumping topics a little bit, you started the Journal of Machine Learning
Research and served as its editor-in-chief. How did the publication come about? And what do you
think about the current publishing model space in machine learning, artificial intelligence? -
Okay, good. So, it came about because there was a journal called Machine Learning, which still
exists, which was owned by Cluer. And I was on the editorial board and we used to have these
meetings annually where we would complain to Cluer that it was too expensive for the libraries and
that people couldn't publish. And we would really like to have some kind of relief on those fronts
and they would always sympathize, but not do anything. So, we just decided to make a new
journal. And there was the Journal of AI Research, which was on the same model, which had been in
existence for maybe five years or so, and it was going on pretty well. So, we just made a new
journal. I mean, I don't know, I guess it was work, but it wasn't that hard. So, basically,
the editorial board, probably 75% of the editorial board of machine learning resigned and
we founded this new journal. - But it was sort of, it was more open. - Yeah, right. So, it's
completely open. It's open access. Actually, I had a postdoc, George Kanidaris, who wanted to
call these journals free for all. Because there were, I mean, it both has no page charges and has
no access restrictions. And the reason, and so lots of people, I mean, there were people who
were mad about the existence of this journal who thought it was a fraud or something. It would be
impossible, they said, to run a journal like this with basically, I mean, for a long time, I didn't
even have a bank account. I paid for the lawyer to incorporate and the IP address, and it just did
to cost a couple hundred dollars a year to run. It's a little bit more now, but not that much
more. But that's because I think computer scientists are competent and autonomous in a way
that many scientists in other fields aren't. I mean, at doing these kinds of things. We already
types out our own papers. We all have students and people who can hack a website together in
the afternoon. So, the infrastructure for us was like, not a problem. But for other people in other
fields, it's a harder thing to do. - Yeah, and this kind of open access journal is nevertheless
one of the most prestigious journals. So, it's not like a prestige and it can be achieved without
any of the- - Paper is not required for prestige,
it turns out. - So, on the review process side,
actually a long time ago, I don't remember when, but I reviewed a paper where you were also a
reviewer and I remember reading your review being influenced by it. It was really well written. It
influenced how I write feature reviews. You disagreed with me actually. And you made it
my review, but much better. But nevertheless, the review process has its flaws.
And how do you think, what do you think works well? How can it be improved?
- So, actually when I started JamLR, I wanted to do something completely different.
And I didn't because it felt like we needed a traditional journal of record. And so, we just
made JamLR be almost like a normal journal, except for the open access parts of it, basically.
Increasingly, of course, publication is not even a sensible word. You can publish something by
putting it in archive so I can publish everything tomorrow. So, making stuff public is, there's no
barrier. We still need curation and evaluation. I don't have time to read all of archive.
And you could argue that kind of social thumbs upping of articles suffices, right? You might say,
"Oh, heck with this. We don't need journals at all. We'll put everything on archive and people
will upvote and downvote the articles and then your CV will say, "Oh man, he got a lot of upvotes."
So, that's good. But I think there's still value in careful reading and commentary of things. And
it's hard to tell when people are upvoting and downvoting or arguing about your paper on Twitter
and Reddit, whether they know what they're talking about, right? So, then I have the
second order problem of trying to decide whose opinions I should value and such. So, I don't know.
If I had infinite time, which I don't, and I'm not going to do this because I really want to make
robots work, but if I felt inclined to do something more in the publication direction,
I would do this other thing, which I thought about doing the first time, which is to get together
some set of people whose opinions I value and who are pretty articulate. And I guess we would be
public, although we could be private, I'm not sure. And we would review papers. We wouldn't
publish them and you wouldn't submit them. We would just find papers and we would write
reviews and we would make those reviews public. And maybe if you, you know, so we're Leslie's
friends who review papers and maybe eventually if we, our opinion was sufficiently valued,
like the opinion of JMR is valued, then you'd say on your CV that Leslie's friends gave my paper a
five-star reading and that would be just as good as saying I got it accepted into this journal.
So, I think we should have good public commentary and organize it in some way,
but I don't really know how to do it. It's interesting times.
- The way you describe it actually is really interesting. I mean, we do it for movies,
imdb.com. There's experts, critics come in, they write reviews, but there's also
regular non-critics. Humans write reviews and they're separated.
- I like open review. The iClear process I think is interesting.
- It's a step in the right direction, but it's still not as compelling as
reviewing movies or video games. I mean, it sometimes almost, it might be silly,
at least from my perspective to say, but it boils down to the user interface,
how fun and easy it is to actually perform the reviews, how efficient, how much you as a reviewer
get street cred for being a good reviewer. Those human elements come into play.
- No, it's a big investment to do a good review of a paper and the flood of papers is out of control.
Right, so, you know, there aren't 3,000 new, I don't know how many new movies are there in a year.
I don't know, but there's probably gonna be less than how many machine learning papers
are in a year now. And I'm worried, you know, I, right, so I'm like an old person, so of course,
I'm gonna say, "Rawr, rawr, rawr, things are moving too fast. I'm a stick in the mud."
So I can say that, but my particular flavor of that is, I think the horizon for researchers
has gotten very short. That students want to publish a lot of papers and there's a huge,
there's value, it's exciting and there's value in that and you get patted on the head for it
and so on. But, and some of that is fine, but I'm worried that we're driving out people who
would spend two years thinking about something. Back in my day, when we worked on our theses,
we did not publish papers. You did your thesis for years. You picked a hard problem and then
you worked and chewed on it and did stuff and wasted time and for a long time. And when it was,
roughly when it was done, you would write papers. And so I don't know how to, and I don't think
that everybody has to work in that mode, but I think there's some problems that are hard enough
that it's important to have a longer research horizon and I'm worried that
we don't incentivize that at all at this point. - In this current structure.
- Right. - Yeah. So what do you see
as, what are your hopes and fears about the future of AI and continuing on this theme? So AI has gone
through a few winters, ups and downs. Do you see another winter of AI coming? Are you more hopeful
about making robots work, as you said? - I think the cycles are inevitable,
but I think each time we get higher, right? I mean, so, you know, it's like climbing some kind
of landscape with a noisy optimizer. So it's clear that the, you know, the deep learning stuff has
made deep and important improvements. And so the high watermark is now higher. There's no question,
but of course, I think people are overselling and eventually investors, I guess, and other people
look around and say, well, you're not quite delivering on this grand claim and that wild
hypothesis. It's like, probably it's going to crash some amount and then it's okay. I mean,
but I don't, I can't imagine that there's like some awesome monotonic improvement from here to
human level AI. - So in, you know, I have to ask this question. I probably anticipate answers,
the answers, but do you have a worry, short term or long term about the existential threats of AI
and maybe short term, less existential, but more robots taking away jobs?
- Well, actually, let me talk a little bit about utility. Actually, I had an interesting
conversation with some military ethicists who wanted to talk to me about autonomous weapons.
And they were interesting, smart, well-educated guys who didn't know too much about AI or machine
learning. And the first question they asked me was, has your robot ever done something you didn't
expect? And I like burst out laughing because anybody who's ever done something on the robot,
right, knows that they don't do much. And what I realized was that their model of how we program a
robot was completely wrong. Their model of how we can program a robot was like Lego Mindstorms,
like, oh, go forward a meter, turn left, take a picture, do this, do that. And so if you have
that model of programming, then it's true. It's kind of weird that your robot would do something
that you didn't anticipate. But the fact is, and actually, so now this is my new educational
mission. If I have to talk to non-experts, I try to teach them the idea that we don't operate,
we operate at least one or maybe many levels of abstraction about that. And we say, oh,
here's a hypothesis class. Maybe it's a space of plans, or maybe it's a space of classifiers or
whatever, but there's some set of answers and an objective function. And then we work on some
optimization method that tries to optimize a solution in that class. And we don't know what
solution is going to come out. So I think it's important to communicate that. So I mean, of
course, probably people who listen to this, they know that lesson. But I think it's really critical
to communicate that lesson. And then lots of people are now talking about the value alignment
problem. So you want to be sure as robots or software systems get more competent, that their
objectives are aligned with your objectives, or that our objectives are compatible in some way,
or we have a good way of mediating when they have different objectives. And so I think it is
important to start thinking in terms, like, you don't have to be freaked out by the robot apocalypse
to accept that it's important to think about objective functions of value alignment.
>> Yes. >> And that you have to really,
everyone who's done optimization knows that you have to be careful what you wish for, that,
you know, sometimes you get the optimal solution, and you realize, man, that objective was wrong.
So pragmatically, in the shortish term, it seems to me that those are really interesting and
critical questions. And the idea that we're going to go from being people who engineer algorithms
to being people who engineer objective functions, I think that's definitely going to happen. And
that's going to change our thinking and methodology. >> You started at Stanford
philosophy, that's where you should go back to philosophy. >> Philosophy, maybe.
>> Designing objective functions. >> Well, I mean, they're mixed together,
because as we also know, as machine learning people, right, when you design, in fact, this is
the lecture I gave in class today, when you design an objective function, you have to wear both hats.
There's the hat that says, what do I want? And there's the hat that says, but I know what my
optimizer can do to some degree. And I have to take that into account. So it's always a tradeoff,
and we have to kind of be mindful of that. The part about taking people's jobs, I understand
that that's important. I don't understand sociology or economics or people very well.
So I don't know how to think about that. >> So that's, yeah, so there might be a
sociological aspect there, the economic aspect that's very difficult to think about. Okay.
§04Transcript · 58:56–1:01:04
>> I mean, I think other people should be thinking about it, but I'm just, that's not my strength.
>> So what do you think is the most exciting area of research in the short term,
for the community and for yourself? >> Well, so, I mean, there's this story
I've been telling about how to engineer intelligent robots, right? 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. And the question is, what's
the most effective strategy? And we've tried, and there's a bunch of different things you could do
at the extremes, right? One super extreme is we do introspection and we write a program. Okay,
that has not worked out very well. Another extreme is we take a giant bunch of neural
goo and we try to train it up to do something. I don't think that's going to work either.
So the question is, what's the middle ground? And again, this isn't a theological question
or anything like that. It's just like, how do we, what's the best way to make this work out?
And I think it's clear, it's a combination of learning. To me, it's clear. It's a combination
of learning and not learning. And what should that combination be? And what's the stuff we
build in? So to me, that's the most compelling question. >> And when you say engineer robots,
you mean engineering systems that work in the real world? Is that, that's the emphasis?
Last question. Which robots or robot is your favorite from science fiction?
So you can go with Star Wars or RTD2, or you can go with more modern, maybe Hal from-
>> No, I don't think I have a favorite robot from science fiction.
>> This is back to, you like to make robots work in the real world here, not in-
>> I mean, I love the process. And I care more about the process.
>> The engineering process. >> Yeah. I mean, I do research because it's fun,
not because I care about what we produce. >> Well, that's a beautiful note, actually,
to end on. Leslie, thank you so much for talking today.
>> Sure, it's been fun.
[NO SPEECH]