Placeholder Image

Subtitles section Play video

  • I want to thank the organizers for inviting me to this.

  • This is way outside my usual area.

  • I'm a mathematician, and the closest I get would be CIDCOM,

  • but this has been a lot of fun.

  • And I have this incredibly pretentious title.

  • And so I'm going to try to explain

  • to you what I mean by this.

  • Online I have a bunch of videos that go

  • into this in a lot more detail.

  • So sort of think of this as a quick preview of the videos.

  • And I have a lot of people to thank,

  • not enough time to give them all the credit they deserve.

  • So what I'm interested in is these sort of major transitions

  • in evolution.

  • But they're also changes in architecture,

  • and you see increases in complexity and elaboration

  • of networks.

  • Unfortunately, those are the four most confused subjects

  • in all of science.

  • And engineers know a lot about these things,

  • but they keep that to themselves.

  • So I'm going to focus on two parts of this.

  • Of course, you're interested at this stuff at the top.

  • But I'm going to kind of use bipedalism, that transition,

  • as an example.

  • If we hadn't done that, none of the rest would have happened,

  • so that's a really crucial--

  • they're all crucial, but that one's particularly crucial.

  • And how do I explain universals?

  • Well, my main way of doing it is with math.

  • But we'll not do that today.

  • We'll focus on trying to look at some diverse domains, so not

  • just networking, but like I said, bipedalism and our brains

  • and how our brains work.

  • Currently, unfortunately, we have

  • kind of a fragmented theory behind this.

  • And so one of the objectives of my research

  • is to try to get this to be not a whole nine

  • subjects, but really one.

  • And that's the framework to try to do this,

  • is to create a theory which can then

  • be used to understand these transitions.

  • And again, lots of details in the videos.

  • So now, I'm very different from this community.

  • Maybe only one letter different, but that

  • makes a big difference.

  • But I think there's a lot of things

  • that we're also interested in common.

  • We want to have all these features.

  • I may be more theoretical.

  • Maybe you're more practical.

  • But I think we also, again, maybe

  • have different priorities but the same interests.

  • And also dynamic and deterministic.

  • And by deterministic I just mean in the way

  • I think about the problems today,

  • I focus on not average behavior, but kind of what

  • goes on worst case.

  • And so in bipedalism, one of the most important things

  • is a trade-off between robustness and efficiency.

  • Now of course, we'd like to be both.

  • We'd like to be in the lower left hand corner.

  • That's the ideal case.

  • And if you compare us with chimps, for example,

  • at distance running we're about four times

  • as efficient as they are, and that's really substantial.

  • And if you've got a bicycle, you get another factor

  • of two or so, roughly, again, roughly.

  • But much more fragile.

  • And the bike makes the crashes worse,

  • and so that's the trade-off we see in adopting bipedalism.

  • And so what I want to do is think about these kinds

  • of trade-offs.

  • We'd like to be cheap, we'd like to be robust.

  • But it's hard to be both.

  • Now the cardiovascular physiology part of it

  • is very interesting as well.

  • We have a very upgraded cardiovascular system

  • compared to chimps.

  • If you want to read about that, that's a recent paper.

  • And I have some, again, videos online on this physiology.

  • So we'll not talk about physiology.

  • We're going to worry about the balance part of it,

  • and not worry about efficiency, but really robustness.

  • And ideally, again, we'd be cheap, fast, flexible,

  • and accurate.

  • We'd have all these things.

  • Again, I'm going to ignore the cheap dimension.

  • PowerPoint only lets you really draw things in two dimensions,

  • so we're going to keep projecting things

  • into two dimensions.

  • So again, we'd like to be fast, flexible, and accurate,

  • but it's hard to be all of those things.

  • So what I want to talk about is the trade-off

  • in layered architectures, and focus

  • on a very simplified view of what our brains do,

  • which is planning and reflexes.

  • And as an example, this task.

  • This is not me.

  • I'm more of an uphill kind of guy.

  • So if this is me, we'd be watching a crash.

  • But what we can see here is this higher level

  • planning using vision is slow but very accurate.

  • And then you have a lower level at the same time,

  • a reflex layer, which is fast dealing with the bumps.

  • So you've got the trail you're following and the bumps.

  • And so we can think about this planning layer.

  • It's slow, but it gives us a lot of accuracy, flexibility,

  • it's centralized.

  • It's conscious, deliberate.

  • And it deals with stable virtual dynamics.

  • But just the opposite of the reflex layer,

  • which deals with the bumps.

  • It's fast, but it's inaccurate, rigid.

  • It's very localized and distributed,

  • and it's all unconscious and automatic.

  • And it deals with the unstable real dynamics to create that.

  • So these are really opposite, completely

  • opposite functions that the same nervous system multiplexes

  • very effectively.

  • And so we put those two things together.

  • We're not ideal in the corner, but we

  • behave almost as if we are.

  • And so of course we'd like to be better, faster, cheaper.

  • You can usually choose two or one at best.

  • And again, we're going to focus on this trade-off between fast,

  • accurate, and flexible.

  • And again, projecting very high dimensions into these.

  • And we're going to focus on just these aspects right now.

  • And again, how do we talk about that?

  • Well, again, we have a math framework for that,

  • but I'm going to talk about how this cuts across many domains.

  • So I claim that this is a feature universal, laws

  • and architectures.

  • And again what I mean by law is a law says,

  • this is what's possible.

  • Now in this context, this is what we can build out

  • of spiking neuron hardware.

  • But what is an architecture?

  • Architecture is being able to do whatever is lawful.

  • So a good architecture lets you do what's possible.

  • And that's what I mean by universal laws

  • and architectures.

  • What I claim is, in this sort of space of smart systems,

  • we see convergence in both the laws and the architectures.

  • And so, again, I want to try to talk

  • about this kind of picture, but in some diverse domains.

  • So what are some of the other architectures

  • that look like this?

  • Well, one that you're obviously familiar with

  • is this one from computing, where

  • we have apps sitting on hardware mediated by an operating

  • system.

  • We don't yet really understand what the operating system

  • is in the case of the brain.

  • We know it's got to be there, and we

  • know it's got to be really important,

  • but we're a little murky on it and exactly how it works.

  • So one of the things that I'm interested in

  • is kind of reverse engineering that system.

  • So you're very familiar with the universal trade-offs

  • you have here.

  • So for example, if you need absolute the fastest

  • functionality, then you need special purpose hardware.

  • But you get the greatest flexibility

  • by having diverse application software.

  • But that would tend to be slower,

  • and you've got that trade-off.

  • And then Moore's Law, of course, shifts this whole curve down

  • as you make progress.

  • Unfortunately, there's currently no Moore's Law

  • for spiking neurons.

  • We're kind of stuck with the hardware we have.

  • But the operating system is crucial in tying these two

  • things together.

  • So now we have a computer science theory

  • that more or less formalizes some aspects of this in a sense

  • that if you want to be really fast,

  • you have to have a very constrained problem,

  • say, in Class P. But if you want to be

  • very general in the kind of problems you solve, then

  • unfortunately your algorithms are necessarily

  • going to run slower.

  • It turns out, at the cell level we have

  • the same sort of trade-offs.

  • If you want to be fast, you better

  • have the proteins up and running,

  • but your greatest flexibility is in gene regulation

  • and also gene swapping.

  • So what we have here is these convergent architectures,

  • the fundamental trade-off being that you'd

  • like to have low latency, you'd like to be fast,

  • you'd like to be extremely accurate.

  • But the hardware that we have available to us

  • doesn't let us do those simultaneously,

  • and there's a trade-off.

  • And then we exploit that trade-off

  • to do the best we can with good architectures.

  • So I want to talk a little bit more

  • about this in a little more detail.

  • And I want to kind of go through and use this example of balance

  • as a way of seeing a little bit more detail about how

  • these systems work.

  • And I want to sort of connect the performance

  • with the underlying physiology a little bit.

  • So what we're going to do is we're

  • going to do a little experiment using your brains.

  • And so one thing things I want to do is use vision.

  • And I want you to try to read these texts as they move.

  • So it turns out, up to two Hertz you don't have

  • too much trouble doing this.

  • But between two and three Hertz, it gets pretty blurry.

  • And that's because that's as fast as your eye can move

  • to track these moving letters.

  • Now I want you to do a second experiment, which

  • is to shake your head no as fast as you can

  • while you're looking at this.

  • Now I don't mean big, I mean really fast, OK?

  • And it turns out no matter how fast you shake your head,

  • you can still read, certainly, the upper left.

  • So it turns out your ability to deal with head motion is much

  • faster than for object motion.

  • So why is that?

  • So first of all, evolutionarily why,

  • and then mechanistically why.

  • So there's a trade-off.

  • Object motion is flexible and very accurate, but slow,

  • whereas head motion is fast but relatively inflexible.

  • We'll see why that is.

  • So why is that?

  • Well, when you do object motion tracking, you're using vision.

  • Shouldn't be surprised by that.

  • So vision is very high bandwidth, but it's slow,

  • several hundred milliseconds of delay.

  • That's why you get this two to three Hertz bandwidth.

  • So slow but very flexible.

  • So your visual system did not evolve

  • to look at PowerPoint slides, yet you're

  • sitting here doing that.

  • And it's also very accurate, and we'll see in a minute

  • why the accuracy is there.

  • For head motion, you have a completely separate system that

  • doesn't use vision directly.

  • It has this sort of rate gyros in your ear.

  • And it's very fast, but as we'll see, inflexible and inaccurate.

  • And this is the vestibular ocular reflex.

  • So that's a very low delay system.

  • In