Numenta Brains@Bay, 11 May 2020

Previous Meeting Next Meeting

⏯

youtube image

►

From YouTube: Brains@Bay Meetup - Predictive Processing in Brains and Machines (May 11, 2020)

Description

To learn more about Brains@Bay, visit our Meetup page https://www.meetup.com/BraIns-Bay

Brains@Bay Meetups are designed to bring together the fields of neuroscience and artificial intelligence. Speakers for this meetup have been selected from each discipline to provide unique views of the topic of Predictive Processing. The speakers for this event are Avi Pfeffer & Georg Keller.

Modern neuroscience has gained an increasing appreciation that predicting aspects of the environment, like the sensory consequences of action, is a core component of intelligent computation. In recent decades many models, such as Hierarchical Temporal Memory, have surfaced to propose how the brain generates predictions. In this event, our speakers will talk about one family of such models: predictive processing.

Predictive processing states that the brain uses internal models of the world to continually make predictions about events in the environment. These predictions are compared with sensory input to generate prediction errors, which can be used to update the brain's internal models.

0:00 Georg Keller - Predictive Processing in the Neocortex
42:01 Avi Pfeffer - Building Long-Lived AI Systems Using Predictive Processing
1:17:40 Q&A/ Discussion

A

Hi, everyone welcome to the brains at Payoneer app and it's our first online edition and we'll be talking about predictive processing, brains and machines, and these events organized by Aris, Fisher, Matt, Taylor, myself and Tsutaya Matt.

A

So 28 days, machine learning meets neuroscience meetup we usually meet and Bay Area California, and our goal is to discuss brain inspired, machine learning, algorithms and from both the perspective of neuroscience and machine learning researchers. So we always aim to have at least two talks, one from the neuroscientist and one from machine learning. We try to make these two talks convey interesting discussion at the end, so we previously had the tabs on continuous learning on sparse representations and unhappy learning.

A

And if you have suggestions for the next topic, you can join our meetup page and in there's a message board with topic for suggestions for the next Meetup. So our meetup sponsored by a new mentor who paid for the food and the place, at least in the previous editions. Now it's online, so we don't need it and University of California Santa Cruz was also partnered in previous editions.

B

Right, so we have two speakers for today's meetup. We in a few minutes your Keller, we talked about predictive processing in the cortex and then, after after him, I'll be Trevor, we'll talk about a new system, they're developing for building long-lived AI systems with inspired by predictive processing and after world afterwards, we'll have 20 minutes and maybe more if time permits for some Q&A.

B

So with this, it's my pleasure to introduce Gary Keller Gary was actually my PhD supervisor. Jürg work guaranteed his PhD at the ETH Zurich, where he worked on the feedback processing of Sam birdsong vocalizations. He later moved to the Max Planck Institute for neurobiology in Munich, where he found sensory motor mismatched signals in behaving mice, and this is sort of some of these works- are some of his first steps into working on predictive processing. So, since 2012 he has a lab at the fmi in basel, where the lab mostly works on testing.

B

The hypothesis of whether predictive processing is a canonical cortical computation. So without further ado, uh your girlfriend.

C

As a general introduction, I try more recently to start from a point of view that should be familiar to most of you, whether in your in machine learning or in neuroscience, and that's the idea of reward prediction errors, a signal that was discovered a few decades ago- that that in essence gives you a signal of how how much of a violation something is from the reward you're expecting to get and in one of the classical models for describing the word prediction errors.

C

Or you know the theoretical backdrop behind the concept is TD learning temporal difference, learning where you have a state value B here in this case and a word prediction, error, Delta, and that this is a stolen from a paper of map, often acknowledged either that that's about a variational TD learning that I will not get into. But they have a nice illustration of this. This idea of the state variable in a reward for dictionary where the reward detection error can be used to update the state value.

C

Now, the the usefulness of this is that you can do other things with reward. Sorry with the reward prediction error as well, and that is update your static weight, matrix and and that's why, in such a useful idea, now so sort of developing from that I'd like to explain what I mean with a predictive processing framework for cortex, and so what I've done here is used the same type of representation where there's a state variable at the top and reward prediction area.

C

I mean this is just the identical scheme now to sort of illustrate that the counterpoint or the thing I'm trying to distinguish this model or this category your family of models from the you know when you think of cortex, even to this day, there's this idea of feature detectors as they're called in neuroscience. The idea that you, you know, build cortical representations through a a bank of filters and and obviously I mean this isn't isn't wrong.

C

But the schematic idea for this sort of to highlight the differences is that you have internal representations, that sort of build upon each other, and that's also how we describe neural networks and usually there's the sample arm of a hierarchy. On top of this, you know where you start with simple representations, and then you know build your corner triangle or cupcake cells from more simple representations and the higher you go in this higher get more complex. These representations become, but again this is just. Obviously you can do this as well with a pretty processing model.

C

Now, just for Naaman cloture, we call this an internal representation or the in the model. This is called an internalization of the state variable and the reward prediction error becomes a prediction error, but now you can string these modules. If you will these computational modules together, you know to form your favorite hierarchy and and again this is what we call a predictive processing framework and, as opposed to world, is more classically described as a representation.

C

For me, this is the term David Moore coined in his description of cortical function, and but you know one way to think about. This is a hierarchy indeed, and one very popular one, but as I'll try to allude to at the end, there's other ways of thinking about this. This doesn't need to be a hierarchy.

C

The ideas that you have computational units that can be combined into a network in a in a relatively arbitrary fashion, but the important thing here is that the predictive processing framework I use that term loosely to describe a computational algorithm that has an explicit representation of a prediction error all right.

C

So the difference between this type of model and this type of model is that you have an explicit representation, meaning a neuron that codes for the prediction error explicitly with its Viking right, then sort of to not skip over the computational background that this, the very influential early work on the concept was: it was a paper by around Ballard in 1999, were they I think we're the first to describe the the computational framework? For this sorry I should have mentioned this. This was largely based on evidence of n stopping.

C

This is a phenomenon that was well known at the time. It's in visual cortex, where, if you show a short bar stimulus like this to a cell and it's aligned perfectly the receptive field of that cell, you get a much stronger response than if you extend that bar past.

C

The classical receptive field of the same cell- and this was known as n, stopping like the the response is larger when you have a shorter bar and and what the round Ballard at the time argued is that one way to describe this is as an error signal, because if you assume that this is the signal, as you get as a prediction from a higher level in the hierarchy, there's a violation of this prediction here.

C

If you only see a short bar and and and they formulated this idea of again in a hierarchical fashion, where we send a feed-forward error signal that is computed, there's a comparison between the feed-forward signal from the layer below and a top-down prediction, and then that was their formalism to describe this phenomenon of in stopping now.

C

What's the evidence that we have accumulated students for for this idea- and am important to note here- is that one of the one of the key problems with with describing this framework is that if we go back briefly to this schematic, it's very simple to experimentally control the input, it's very, very difficult to experimentally, control or record.

C

Even what the internal prediction of the model may be, unless you are willing to actually record from the axons that send these predictions, so so, what's the evidence and the evidence just to highlight again, you know what was the basis of the reward prediction error idea that was the discovery of a neuron.

C

This is again from the 90s that signals an excess of predicted reward with an increase in firing rate and a absence of a predicted reward with decrease I'm sure most of you will be familiar with this with this idea, but you have this bi-directional signal that was discovered. That would signal more or less reward than anticipated. Now again in in cortex, we have this problem of it not being entirely trivial to control the expectation of the animal right in these experiments.

C

The argument is that, because the animal has seen this reward hundreds of times, it has now learned to expect the reward, and the absence of the expected reward is a violation in visual cortex. We need to we use a proxy for this, and this is usually movement. Rightfully stop this. um So the argument is as follows: as an animal grows up in the world, sensory feedback is usually coupled in a very predictable way to movement.

C

So as the animal, for example, moves forward in the world, there's a visual flow to the back, and the animal has learned that through experience I mean so the argument goes, and so we can use the movement of the animal as a proxy for the what the animal is expecting to see as it's moving forward. It's expecting to see a visual flow to the back, and we can do this experimentally and it's a relatively simple, one-dimensional coupling of forward locomotion backward a visual flow and we can violate the coupling by simply halting the visual flow.

C

It's in a very easy, a paradigm experimentally and then can record what happens in visual cortex in primary visual cortex that the area processing, sensory information, the cortex, a visual information in accordance when we perturb this feedback. When there's a mismatch between what the mouse to presumably expect to see and what it actually season, because it's visual. It's because it's a sensory area we can always do is just what we call it.

C

Passive control, an open-loop scenario where the mouse just passively use the same stimulus, and this allows us to isolate which components of the signal are actually violation of a spectating given movement and and which are just explained by the sensory stimulus alone and.

C

While doing this, we record neural activities, I had left this interest for those of you, who've never seen this. This is two photos on calcium imaging here from labor to three of primary visual cortex, and then we can si SS so that the fluorescence here corresponds for the activity of one of these cells and and we can measure how do these cells respond to this violation of expectation, and- and this is what you're seeing here- this was very early finding that we had that there's.

C

When you have these violations, we call them feedback mismatch and layer. Two three on average responds quite strongly. This is the average response of all the neurons that we recorded at the time in and layer two three-and and just to highlight this once if you expose the animal passively to the same stimulus, so it just sees the same stimulus.

C

You get very little to no activation in later two three so most so this difference between you know: they're, both identical visual stimulus except here itself generated, and it's just break and coupling between what the animal expects to see we think and what it actually sees, and so this is what we call a feedback, mismatch signal or now there's a negative prediction: error signal and sort of as a second piece of data that this is specific to visual cortex. What you're, seeing here is is Whitefield, calcium imaging.

C

This is a raw movie of how this looks if you just express one of those genetically encoded, calcium indicators, brain wide and then just look at the activity patterns across the entire dorsal source. Sorry, across the entire dorsal surface of the cortex, those of you not familiar with Mouse and that to me this is the visual cortex, the part that process visual information. This would be somatosensory, cortex, orbitofrontal, entry, single at etc. So it's it's. It stands from the posterior to the most anterior parts of coordinates, and now we can trigger the activity this.

C

This raw activity on these mismatched events and see whether we actually indeed find a signal, that's specific to primary visual cortex and, as I explained before, I'll stop with you briefly. This is primary visual cortex and the average responds to these feedback. Mismatch events seems to be centred on primary visual cortex, so we think I mean based on this and other evidence. We think this is a signal. That's computed locally in this area we're starting.

C

So we found something a way that we think is a prediction error with which we can drive activity in visual cortex. So a question I often get at this point is whether this is a mouse thing, and this is often arose. I.

C

Anticipate the first evidence actually, for you, know, prediction error: signals in primary: visual cortex was human work from 2007, where the experimenters coupled in participants is fMRI a virtual hand to actual hand, movements and then broke that coupling, but there's similar signals we discovered in primary auditory, cortex and normal sets. We found them in bird pallium and then later in in the mouse.

C

The signals I just described to you now important to note is that the first three variants of doing this experiment always were a combination of an absence of an expected signal and the addition of an unexpected one. So it was, it was always two signals at the same time in the auditory cortex, it's easy to understand, because that's it's very hard to to temporarily silence the sensory input to self-generated vocalizations. Now in visual cortex. We have this advantage that we can.

C

We can do with the experiment, one directional, where we just take away and expect the dim, but we still skip these signals. We also call this. This is the origin of the idea of a negative prediction error. It's the absence of an expected signal, whereas a positive prediction error would be the appearance of an unexpected signal. Alright,.

D

So question there: yes yeah, so yes, that's really interesting. So it sounds like you're seeing the mismatched signals, not just across species but across multiple sensory modalities. As well said, you know, potentially a very general property. Have you do you know of evidence in sort of higher level areas as well of any sort of mismatched signals.

C

This short answer is: there's a lot of signals that would be consistent with prediction errors, they're, also consistent with lots of other things. The problem with identifying whether it's a prediction is that you need to know what the coding space of the area is right here. It's it's relatively simple. If we rely on the fact that we think visual cortex codes for visual stimuli index, we know what a visual prediction error would be, but I guess the short answer is a reward.

C

Prediction error is a very, very high level version of the prediction error to reward, and you know signals like reward prediction: errors that there I'm gonna go out on a limb here and argue that they're probably not computed in the dopaminergic system. This right that the expectation for when the monkey should receive a reward is probably not computed in the dopaminergic system and is inherit somewhere from cortex or striatum.

C

So I would argue there are signals that are consistent in higher areas. With a caveat that this, you know without knowing what the codings space of the area is, that you're looking at is very hard to make sure that these are reward code dictionary. Thank you. That's a satisfactory answer. Well, that's great! Thank you.

C

Yes, so so a one-way right that there's many there's an infinite number of models that would technically be consistent. You know that would explain these prediction error signals. The simplest version is just a subtraction I should I should emphasize. There are many other ways and variants that you can achieve this type of response, but I would argue the simplest one is just a subtraction of two signals right.

C

If you have a bottom-up signal, it tells you what you're actually seeing and you pop down signal that it tells you what you're you know what you predict to see. You can just compute the difference between these two signals and you have a prediction error. Now, there's a in cortex! There is a slight problem with this. If I, you know go back briefly to the reward prediction error neurons here at the dopaminergic reward prediction area: knurls.

C

What you see is that there's an increase of firing rate to the to the appearance of an unexpected reward and it decreased the firing rate to the absence of an unexpected sorry to the absence of an expected reward. So, in essence, is known in in theory, could signal the reward prediction error by directionally increases, meaning more decreases meaningless. Now, in cortex the firing rate to these layer, two three neurons is close to zero or it's you know extremely low I've said such that the solution, cortex or I would argue.

C

The solution that cortex is found is just to split the computation across two populations of neurons.

C

You know that because this is bounded at zero right if s is larger than penis neuron, what we will just be silent, but one thing you can do is just flip it around and just do the subtraction the other way, and now you have a sort of a push-pull mechanism. One neuron is active. When there's you know, top-down input is stronger. The other one is active from the bottom-up, but is stronger I. Think there's, there's good computational reasons to do this.

C

One of them is that now your what you're optimizing or your learning rate can just minimize the firing rate to these two populations of neurons, because they're at zero. Anyway, you can just change your synapses, such that both of these neurons are silent right. So what I'd like to do briefly is show you some more recent evidence that this might actually be the case that there's two populations of cells that interact and this push-pull kind of way with the top-down prediction and the bottom-up sensory input, and to do this. This is work by Rebecca Jordan.

C

What we would what Becky is done is record intracellular from neurons in Larry, two three and layer, five in this paradigm, that I just introduced with with the entire feedback, mismatch and playback stimuli, to see whether there are indeed layer, two three neurons that that are in this push-pull situation, that get excitatory top-down input, inhibitory, feed, bottom-up input and, and vice versa- and the short answer is that yes, I mean that's that's what Becky finds is that what you're saying here is membrane potentials on these mismatched events right means the neuron.

C

Depolarizes is more excited which can happen either through direct excitatory input or a belief, as are a decrease in the inhibitory input, and then the blue or neurons that hyperpolarize here's the average, maybe potential or are either inhibited or you know, undergo a decrease in excitation, and the interesting question for us is: are the neurons that you know get the are depolarized by mismatch of those, the ones that are hyperpolarized by visual input and vice versa, and that indeed seems to be the case on average that that we have this push-pull there's two categories: it's not I mean, as you can see, it's a continuum but there's two extremes here and if you look at the the two inputs, what you can see is that there's this mutual antagonism of the mismatched signals in the visual input consistent with this idea of open sort of a push-pull, and this is just plotting- the data on the x-axis, the visual flow response on the y-axis, the mismatch response.

C

So this is very nice correlation in layer two three and this is all cells. This is not the sub set of selected cells. Now the question is that Becky set out the dress is whether this is specific to layer two three and the answer is at least in info granular layers. Here we call it layer five six, because we have a hard time telling whether it's truly a lair, five neuron or whether we're already in layer six. So we've grouped their this into a layer.

C

Five, six category where most of the neurons are probably layer five now the first thing is they do most of these neurons actually have these classical visual responses. If you remove the visual flow stimulus, they they hyperpolarize, and even if the ones that have these depolarizing responses, they don't have an inhibitory and a visual flow input.

C

They're also excited by visual flow.

C

So we I mean in a nutshell what we think the evidence is telling us is that layer, five six neurons tend to integrate the two signals that the bottom-up and top-down input, or have responses that would be consistent with some form of integration, whereas in layer two three, you have this difference computation between two signals right, so that's in essence, the evidence we have I think that's that here, I'm done with biology now the key component missing in this model is, is this internal representation, something that maintains the state variable and again here the model is just one variant, but this could be in CA in the model.

C

The idea here is that you have an you call the putative layers five so that integrates over positive and negative prediction. Error signals. Obviously I mean this. Doesn't here it's because we have zero evidence, that's that this is actually how it's done. This is just one way this could be done and something we are testing now to see whether we find evidence for something like this in the compilot, their computational implementations that I'll show, at the end, there's different variants of doing this.

C

This is just one very simple way of implementing this integration and also importantly, it it makes this unit a expandable unit, because these internal representations function like top-down predictions.

C

If you assume you were in a system that only has one area, this is the you know that the base layer of prediction that you implement here and and that's just the actually the original idea of predictive coding that came from video processing- was that a good prediction of the next video frame is the current video frame, and what you want to signal is just differences between the two, and this is this. Is that idea right?

C

This is it's just this neuron represents the the current best estimate of the visual scene, and all you need to compute is the difference between the bottom-up input and your best estimate at the moment to have annotation coding system. So you know you: can you can run this in independent computational units I mean that's the idea of this toy circuit that we chose here then something I should point out for all the neuroanatomists.

C

Obviously this is a slightly different model from what you would find in a textbook, but the important point that I'd like to point out is that if you run experiments in a condition where you have no control over this top-down input, this entire circuit very nicely reduces to the classical model that you know. People have been building for cortex, where you know, there's a bottom-up input that feeds into layer.

C

4 goes to layer, 2 3 and it's then fed to layer of 5, and- and you can see this- that you know if this is 0 or you're, in a condition where you have no control over the. What the animal expects and you know, average over trials. The the response of the neurons will just go to zero and all you're left with whoops is.

C

Is these two responses that so so the predictive processing model quite nicely reduces in the in the case where you have no control over the top time predictions to what we already know from all the work on cortex over the past decades. So to conclude, what's the point? Why why why go through the trouble of explicitly representing prediction errors at every I'm, not going to say every level of the hierarchy, but every in every computational unit?

C

Why include a prediction error computation in every computational unit and there I would say because prediction errors have a very nice dual function or a dual role: I guess, and one we've already seen right in the in Destroy circuit, the prediction errors function to update the internal representations.

C

That's the same is true for the reward prediction error where you know you state without variables, updated by the reward prediction error now, but these reward prediction errors also have a second, very useful role, which is that they allow you to change the weights throughout your network, assuming you have entered training and you can back propagate these errors throughout your network. Now. Is that the case you know here as well? Can you think of this?

C

The second role of these rewards, but they re of these prediction errors to change synaptic weights, and the answer is yes. Oh sorry, I would argue yes absolutely and to illustrate this. Let's just take a toy model, and this is based on on. You know, work from the from the 90s by Jordan and rumelhart and and Daniel walpert than others.

C

Let's assume we have a very simple Brian just propose to loops, so it's my twenty five minute timer, reminding me that I should stop soon so assume you have a toy brain with just two sensory two areas, one a sensory area and one a motor area now, I'm assume also that your interested in imitating and something that you see our imitation learning in this case. This can be, you know, auditory. You hear something you want to figure out what the motor command is that you need to execute to reproduce what you've just heard.

C

But you know the same is true for any sensory motor interaction. Now this is a difficult problem. This is called or what this is, what they called an inverse model and mapping from sensory coordinates to motor coordinates right. You're, certainly capable of this I, can give you any weird sounding Swiss tournament, word and you'll be able to at least speak this somewhat sensible on first trial right, you don't need to practice.

C

You know what the motor commands are, roughly, that you need to execute to reproduce what you just heard ii have a few seconds ago and that's a relatively impressive skill, but- and this is what they called an inverse model this map and and it's difficult to learn, and they argued at the time that one very simple, sorry, one way it's not simple, but one way to learn. This is to first learn and map a forward model right.

C

So you you, you have a map that goes the other direction from motor at the sensory, and why is this easier? Because you have a second pathway that goes through the world, but you can do random things and then measure the sensory consequences and compare those to your predictions of the sensory consequences, and here, quite naturally the prediction error is what you need is a training signal to update your forward model right, so that the prediction error has this dual role: that it can function to update the the map right.

C

The model here is just the synaptic weight matrix that functions to transform motor coordinates, the sensory coordinates and then, as they argue, eloquently a trick to try. This inverse model is now just just go full circle and say you initiate random motor commands takes your forward model once it's learned and then the inverse model should just produce an identity, mapping back to the motor system and you in a similar way. You can then use the motor prediction errors to train your inverse model, but the the more important point is from the sensory motor learning idea.

C

It falls out quite naturally that the the prediction errors are used to update your model. Ironically, this is also where I started my work when we looked at the prediction errors in the songbird. This is precisely why we were looking for prediction. Errors in auditory areas of the song were to see whether we can find any evidence of this sensory prediction error signal that would allow the bird to update its forward knob right. This is a toilet area. I promised you I would say briefly something about the hierarchical aspect of it.

C

Well, with two areas already here, it's very difficult to fit this into a hierarchy. It's relatively symmetric, especially if you start considering that, if you're in visual cortex, you can use all kinds of signals to make predictions about what you're gonna see in a second, but these can be motor signals. These sounds it can be. You know, coding and associated areas.

C

It can be secondary visual areas, as Rowan Ballard argued, but it's probably a combination of a large number of areas, and this makes the idea of a hierarchy relatively complex and nowadays I like to think a bit more as a as a Norn, hierarchical, predictive processing, where areas exchange both predictions and prediction, errors by directionally and sort of the mathematical framework for this, for this idea is it's largely based on on method, cooks, work on the network of relations where you can wire up these areas in a network without needing a hierarchy, and obviously the hierarchy right.

C

If you, if you choose to focus on one area, say visual cortex, any network will look like a hierarchy from the point of view of one of the areas that you randomly select, but the whole network is whole I would argue, is not necessarily oh, it doesn't need to be an arcade very likely in cortex is not we're strictly hierarchy again right. This is not saying that you can't define a functional hierarchy from the point of visual cortex, but that's just one arbitrary choice now does this work?

C

Can you take this predictive processing network and you know implemented one to one and get something useful again? This is not my work, but you know work of David Cox and then ruffle Bolger share it. You can so if you, if you take the exact formulation or almost exact formulation, that Rob Ballard had you can you can build video prediction models that work decently well of just a few weeks ago. A new version of this came out that outperforms this by quite a bit, they'll show in a second um and then maybe more.

C

Interestingly, an argument throughout for Bogart very eloquently made is that if you have this exclusive computation of a prediction error at every level of every in every processing unit, then you can approximate it under certain circumstances. You can approximate backpropagation, so you don't need to back prop the error through the entire network if you compute the error locally in every computational node.

C

This is just from a few weeks ago, literally a updated version of this predictive coding or pregnant on archive here, and you know: they've improved a lot on the performance of the network, simply by changing how the internal representation is updated and they argue it's closer to the closer to the rowan ballard model and that the pregnant the original brad night that a deviance from this model and that the reason they are now doing much better is because they follow this idea more closely.

C

But this is just to highlight that there's not one way to update this internal representation and this toilet circuit that we proposed might be completely wrong and flawed and in time will tell a lot of our work is now on trying to understand what the whether we can find molecular markers for these cells to give us an experimental handle test, these models more explicitly in Mouse cortex and maybe just to conclude on one final slide of biology. The other reason why I'm very excited about these these models in these ideas is that.

C

So if you assume that you know all of perception is based on this constant comparison between some internal model and sensory input and that there is this push-pull between two types of prediction: errors that update this, you know you can imagine scenarios where there's there's a bias in this computation one way or the other, and- and this might lead to things like hallucinations right, if you're, if you're, not if you're a circuit is, is driven too strongly or determined, or the update into a model update, is driven too strongly by top-down input.

C

You might head to a loose tonight, and so you know, excuse me relatively lately along these lines.

C

What we're trying to see is whether, if we give a mouse entity, does it actually shift this balance and and very short and unfair summary of this work, but if I'm a deal signed- or at least yes, it probably does here- is a you know: prediction error response under antipsychotics clozapine and just some control signal ammo to relate the control signal and what we're seeing relatively consistently is that these antipsychotics affect these prediction: error response relatively selectable in cortex and leave other signals relatively unaffected. Obviously why this is?

C

We have now cool, but it's sort of a nice entry point for us to try to understand and unravel the system and with this. Thank you very much for your attention. These are the people doing all the work I.

D

Think thanks Garrick, that was great I, think we have at least one question from Jeff. So let me try to see if I can unmute him yeah.

E

D

Right yeah, that's great.

E

Presentation, thank you. I was just curious. You know the idea of this internal model, it's something that a lot of people would hear we do, and but I just curious if you thought you'd be trying to like predict frame-to-frame in a video this, but that's that's a sort of a very social simplistic model in general representation, and it's obvious that at least in humans and probably other mammals, there's a very, very rich in complex internal representation of the world. That can't be. You know the certeyn from moment to moment sensory input.

E

So is that an area you think about? Do you think that's just another problem that can be dropped into this system, I just curious. What your thoughts about that are! It's.

C

An interesting question um so all the applications we have, as you point out rightly or there are toy models or toy problems in video frame. Predictions now I mean my I. Don't have any good evidence for this, but I would suspect that you can expand this internal representation in dimensions that are not quite as trivial and as soon as you wire up a network that it's not hierarchical.

C

That has different computational units that that can build a representation across modalities and in probably time scales that you then get much more complex internal representation and, as a consequence, behavior.

E

C

It's a your hope is.

E

It that the coding model could expand to include much richer detail. Yes,.

C

Absolutely so I mean this is this is what I'm I'm and you know I, don't have good evidence for this, but one one you know baked piece of evidence that that gives me hope is that if you are that, if you look at that, you know there's a lot of assumptions here. Let's assume that something like you know, neuropsychiatric disease is actually somewhat related to the name balance in this internal model.

C

What you do see is that both autistic and schizophrenic patients, they have primary sensory deficits and relatively high level of cognitive tests them deficits in the interactions with other people right. So there is at least the idea that that you can use the same model to describe sensory processing as as as well as social interactions. Thanks.

D

Great thanks so I know: does anyone have any anyone else have any quick questions they want to ask? We do have time at the end, I have a couple of high-level questions I'd like to ask, but I'll leave that till the end as well. But if there's any other questions please do, let me know on chat.

D

Okay, so let me unmute Abby who's who's online, welcome, Abby! Thank you very much. That's and Aras. Do you want to introduce Abby.

B

Of course, so avi is the principal scientists at Charles River analytics in Boston. He previously worked with Stuart Russell I believe and then got his PhD from Stanford. He developed the first general-purpose probabilistic programming language, while at Harvard and has generally avi has been a pioneer in probabilistic programming and various types of probabilistic probabilistic reasoning, so I'm very excited to hear his talk and I learned randomly like about a year ago that he's he and his team have been working on the predictive processing.

B

Inspired Network, so I know I'm excited to see what he has to say.

D

F

You thank you really appreciate that introduction. So first question is: does anyone know how to use Google slides because I had to do this on a personal laptop because I couldn't get through on my work? Computer and I don't know how to make it go fullscreen.

F

D

I believe there should be a starch slideshow button or or a present button. Do you see that.

F

It's this one which says explore.

D

Yes and near the top right, there should be a present button. You can see.

G

F

Don't see any present button, you.

A

F

Well, I can make a screen and at least you'll see yeah.

D

It should be under view on there's also an if you look at the menus. There's a you go to the view menu. There should be a present button or present menu item under there.

F

Sorry, maybe I'd be better off using Chrome. That would make sense. Okay, let me try this excuse. It's just gonna take a moment. No.

D

Problem we're all getting used to this.

H

H

H

D

The other thing you could do is you could share your screen already and then we could see what you're doing, and perhaps we could point out yeah just.

F

D

D

H

H

H

H

You you, okay,.

F

Okay and I see present okay.

D

F

Share my screen, then.

D

Yes, we see that okay.

F

D

F

You very much, and thanks for the invitation to talk to you, they're excited to talk to you about some AI connections to predictive processing and just to give you a picture of what we're trying to do we're trying to build long-lived AI systems that sense the environment. They reason they learn, they use some internal representations and then they act on the environment and we're particularly interested in longleaf systems that get better over a long period of time and open world environment where you can't anticipate everything that your AI agent is going to encounter.

F

While you develop that agent, so our goal ultimately is human-level AI and moving towards human level. Ai. There are a number of things that we would need to do. So. First of all, we need to move beyond problem solving to general interactions in open-world environments,.

F

We're talking about AI agents that are constantly learning and evolving and reasoning about a wide range of situations, and they need to take into account the context or whatever situation they're in the more diverse the environment, the broader the set of capabilities needed for the AI agent and transferable knowledge is really vital to enable the agent to operate in diverse environments. Of course, this is a long way off. I'm, not saying we're close to getting human level.

F

A is a long way out away from where we are today, but this is sort of a target that the research that we work is aiming to get towards.

F

So deep reinforcement learning is one approach with very impressive results, but some limitations. It's extremely data hungry, tends to rely on a huge number of simulated interactions to learn.

F

It has unfortunately limited capacity to transfer so far, so one study found for the Atari game, deep reinforcement, learning agents which were one of the successes of the Atari that if you train on a game and then just change the game slightly, it fails miserably. In fact, another story is where a colleague of ours trained an agent to play pac-man and then, if you remove a ghost which theoretically should make the game easier for a human, it certainly would make it easier.

F

The system completely crashes, because it's just remembered how to work in that particular game and I know there are techniques to address this, but it's still an earlier limitation.

F

One issue is that these methods don't really have a notion of causal relationships in the domain. Just correlation doing this makes this hair tend to happen better and also they struggle with open-ended inference. That requires information outside the specific input. So an alternative view is to view neural networks as bottom-up data structures, but if we could combine top-down knowledge with bottom-up sensor processing, maybe we can do better if we have useful knowledge that might make our systems less data hungry general knowledge might be able to help transfer and knowledge can explicitly model causal relationships.

F

Those are some of the benefits and also if we have a representation of knowledge, we could potentially bring open-ended world knowledge to bear, but what we must not lose- and this is really trying to be different from many of the existing knowledge based approaches to AI. We must not lose deep. Learnings ability to learn, complex functions can't be programmed and really rich function.

F

So we want a really in a deep way: try to combine the ability to represent knowledge with deep learnings ability to learn rich functions from data, so people talk about symbolic, sub, symbolic AI, and that's also what I have in mind. I just want to say that there's really two dimensions to the symbolic sub, symbolic synergy that people don't talk about, so one is the form of the model content.

F

Do you have parameters?

F

Do you have vectorize parameters of a transition model, for example, versus functional forms, for example, some sort of logic encoding, the transition model on the other side, and you have the conceptual level ranging from high level knowledge of how the world works, such as scripts, that describe the kinds of interactions that might take place in a given environment, going down to low-level data such as video. So that said, we're taking our in spirit called on a second.

F

Sorry, my children were very loud I, don't know if you could pick that anyway, hey.

D

Army, we have a question from from Jeff Hawkins and the audience. Yes.

E

About your previous life, briefly yeah, so I could use define transition model I'm. Sorry, it may be obvious everyone else. Oh.

F

Yeah, sorry I'm, sorry so on. If you have a dynamic system and you're modeling, the world is a dynamic system. So at each point in time the system is in a certain state and there's some sort of model whatever that ears. That describes how the state of all's from one time point to the next and that could be done by numeric parameters.

F

That could be done in more of a logical way. That says if this, then that alright.

E

Here it you're really referring to more sort of traditional symbolic representations on the right there, like you, know, rules of some sort or something like that right.

F

Or a strips planning system, yeah.

H

F

Sorts of representations would be on the right, whereas on the left it would be maybe a layer in a recurrent neural network or something.

E

So you can are you, including in this transitional object, the idea that most of these transitions might be based on behavior like sensory, you know, I go this way, get one thing that worth noting is that included in that transition model words? Yes,.

F

Certainly yeah all right. Thank you.

F

Okay, so here's where predictive processing comes in here and first of all, my apologies to George and everyone who actually does predictive processing for real because we're really only using it as a metaphor that is sort of guiding the way we think about. Building these models, we're not really doing predictive processing properly, but hopefully you'll. Forgive the liberties that we're taking so I. Don't need to go into this since you've seen a whole talk on it.

F

An but talk about building AI to use predictive processing so so, prior to main knowledge is what we're going to use for making the predictions about the layer below, and then we go to fuse the predictions with errors from the observations and we're going to learn the predictive knowledge from experience. This was interesting to see George's slide on the inverse model on the forward model. Then how you learn the forward model from experience, so that's sort of the idea we have in mind here as well.

F

We also believe in hierarchical representations, where the hierarchy really is describing levels of abstraction of the world, so at the highest levels. You have sort of very general mechanisms for how the world works as you go down the hierarchy of more and more specific and detailed mechanisms for a particular thing in the world: we're modeling temporal dynamics, so this is where the transition models come in and our goal is also to be able to handle surprises. So this comes in with the open world.

F

Types of environments where you're, not you don't know in advance, exactly what you're going to see and so in a predictive processing framework. When you get a surprise of what we try to do is figure out where exactly in the hierarchy, it's surprising, so that we have to update our models.

F

So that's predictive processing, I'm, also going to talk about probabilistic programming, so probabilistic reasoning. So let's take a step back before I talk about probabilists programming. I have to talk about what probabilistic reasoning lets you do it by probabilistic reasoning: I mean you have a probabilistic model of how the world works, how your environment works, the interaction between the agent and the environment and the sensors and the motor capabilities, and you can use that to predict future developments.

F

So if you take a certain action, certain effects will result and the world will transition a certain way. You can also use it to infer past causes that led to current developments, I see something. Why did that happen? Well, if I have a probabilistic model of the world that has causal information, I can say. Oh, something happened to cause that to happen. You can use the past to learn how to predict better as well.

F

F

The challenge with probabilistic reasoning is that if you want to develop a probabilistic reasoning application, you need to develop the representation for that application. We need to develop probabilistic, inference, algorithms and learning algorithms, which is really a lot of work, and for that reason, for many years of probabilistic reasoning, applications were limited to very simple frameworks like bayesian networks, where you just have a fixed set of variables that you're describing.

F

So the goal of probabilistic programming is to drastically reduce the amount of work required to create complex probabilistic reasoning applications.

F

So probabilistic programming provides first of all a general-purpose representation language for models, so you don't have to invent the representation each time and second of all, built-in inference and learning algorithms that apply to models written in the language. So you don't have to write the inference and learning algorithms each time either. So probabilistic programming makes sense or predictive processing. In our view, and again, apologies that I'm not claiming that the brain necessarily uses a probabilistic model to do predictive processing. But it's a it's a natural mathematical framework for predictive processing.

F

So the prediction is the prior probability of something happening so from the layer above it gives a prediction: a prior probability for the layer below would expect to see their errors correspond to likelihoods, which means how likely is the data given what I think might be happening at this level of of the hierarchy?

F

So if the data confirms what I'm seeing is gonna have a high likelihood, if the data is surprising, then what I think's will is happening will have a low likelihood and something else will have a higher likelihood and then the percepts then are the posterior which, by Bayes rule, combines the prior and the likelihood.

F

So why probabilistic programming for predictive processing? Well, it X lets us Express. The prior domain. Knowledge using programs, which are a very expressive way to egg, represent the complex kinds of knowledge that we might have in the world and then alexis using the built-in inference algorithms to take observations, and we concil them with the predictions to form posterior beliefs, and it lets us work with rich data structures and complex processes to handle real-world situations.

F

So taking these two ideas of predictive processing and probabilistic programming, we are developing a framework called scruff, so for scruff is a probabilistic AI framework based on predictive processing and it uses probabilistic programming. But it's more than what previous Pro ballistic programming languages have been it's more than a modeling and an inference language. It's a framework for building AI systems that sense and reason and learn and act in their environment operates on a hierarchical network of variables. I'll show you an example.

F

In a moment, each variable is defined by a probabilistic program that generates a variable at the layer below and what scruff provides is an AI runtime that manages the whole network of variables and all the reasoning that needs to happen for the AI agent to sense and reason about what's going on.

F

So it reasons about the dynamic world, the changing world over time at flexible time granularity and what I mean by that is that it doesn't proceed in lockstep from time, one to time, two to time, 3 and so on. Rather any variable can be instantiated at any time and reasoned about whenever you want. So, if there's a lot of changes happening about a variable, you can reason about it frequently.

F

But if the variable is changing very slowly, you only have to reason about it infrequently and it improves its models over time through three learning mechanisms that I'll describe in a moment, and it's designed to detect and characterize novel situations as well.

F

Okay, so here's an example suppose we're building a system to observe vehicles and track those vehicles and predict what they're going to do. So. This is a hierarchical network that might describe it and I'm just going to look at the right-hand side of this network. So at the bottom level we might have some tracks. Those are even actually, though, those are observed.

F

Tracks of the vehicle now tracks are generated by trajectories, so a trajectory might say: go from point A to point B and as it's going from point A to point B, it's gonna emit certain observations and those observations are going to produce the tracks going up a level. We have a particular vehicle movement command, so a movement command might be drive, East, 10 meters and that movement command generates a particular trajectory through physical space.

F

F

Going up a level we have the movement capabilities of the vehicle, so now we're reasoning at a more abstract level about what what can this vehicle doing general. So, if it's a car, it might be able to move along the axis of its wheels, but it can't move from side to side. Other vehicles might be able to have 360 degree freedom. So this movement capability is a probabilistic model that generates movement commands for the vehicle.

F

Now, what if we have many different vehicles, and we want a reason about new vehicles that might arrive in the scene, so we have a movement capability generator. This generates the capabilities of vehicles. If you have a new kind of vehicle, you say: ok, this vehicle can move like a car. This other vehicle can move sideways. But what happens if we get something really really surprising, like we encounter a vehicle that hops like a bunny?

F

Ok, so for this we have to go all the way to first principle through our movement theory, which might have knowledge of mechanics and ballistics. That would update our movement capability generator to allow for the for the possibility of hopping. So this is how we envision it possible to really extend what this AI system might know about and hypotheses about hypothesis size about when it comes to vehicles.

F

So the inference algorithm we use is called a asynchronous belief, propagation belief, propagation is a classic algorithm for bayesian networks due to perl for non loopy networks. It's an exact algorithm for it's a commonly used approximation for loopy networks and terian teri Ty and I developed asynchronous for the propagation for dynamic situations, with this flexible time granularity and in particular, what you'll find is that those higher levels of the model which describe more abstract concepts?

F

Don't change very often, so we don't read about those very often the lower levels, like the tracks change all the time. So we we reason about them. A lot.

F

So this is what, for example, the hierarchical monitoring using asynchronous belief, propagation might look like with instantiated variables, so we get a track at every time point. The trajectories are also instantiated reasonably often and going up to the hierarchy. Something like the movement theory might only be instantiated once and so within this network that we constantly we constructing and reconfiguring. We use this message passing inference to maintain beliefs about all the variables in the system, so I think.

D

E

D

More question from Jeff all.

E

Right, sorry, be the only one asking questions that please other people turn in I just have a really high level of question about this. This is sort of movement theory you have for vehicles. Is this a programmed or learned movement I mean? Are you actually designing writing code about what vehicles might do when movement do, or is this somehow learned organic?

E

F

That's a great question, so you start out with programmed knowledge, and the movement theory would certainly probably be programmed at this point. However, the learning mechanisms can evolve the representations at different levels, so the movement capability generator might start with a couple of kinds of movement capabilities, but over time you might learn to recognize more different movement capabilities. Do you think.

E

It's possible to build, you know, AI, starting that way, I mean the evidence we would have from humans that almost all the knowledge that high level knowledge, learn and- and the question is, do you think you can get to the human-level AI by starting with this sort of program, knowledge and having the world evolved? Or is this sort of a is the best we can do right now.

F

So I don't know.

F

I, actually, don't know whether that statement is correct. You probably know more than I do whether humans have any innate, not knowledge to physics, for example, that.

E

Wasn't the question you know you think about like vehicles, I, don't think we know of.

F

Course, of course, we don't have an innate knowledge of vehicles.

F

Hey I, don't know how far it's possible to go without any pre-programmed knowledge, it would be nice if it was possible to go all the way. I'm, not I'm, certainly not claiming that you can encode a huge amount of pre-programmed knowledge that would suffice for all the needs of AI. That I certainly do not believe that and that's why learning is very important in this framework, but I do think that a combination of knowledge and learning is probably the right way to go.

F

But, more importantly, knowledge gives you the framework with which you can learn new knowledge that is organized as knowledge. It can help you learn in the future. So it's not so much an argument about knowledge versus of learning as much as it is an argument about having representations that enables you to learn and I. Don't really care that much, whether those representations are human interpretable or not as long as you're learning representations that enable you to learn more.

E

It's to really know yeah.

F

Yeah I think that's bad.

F

Okay, so let's talk about the learning mechanism, so what is Bayesian update, especially for the highest levels of the hierarchy that changed slowly?

F

The observations that we get might help us learn which hypotheses are correct. So, for example, let's say we generate a new vehicle and we're not quite sure what kind of vehicle it is.

F

So we might hypothesize and have different hypotheses one hypothesis is, it can do sideways and another hypothesis is it can't move sideways, and so once we observe it moving sideways, we update our beliefs about those hypotheses to confirm that it does move sideways and we're learning something about that vehicle that can last for a long time, not only there, but we're also learning something about vehicles in general that vehicles are moved sideways or perhaps more frequent than we had previously thought, and so we're learning about the movement generator as well.

F

So the next kind of learning is parameter learning, which is what we're all familiar with from deep nets. And so what we do is we improve the existing hypotheses by adjusting parameters at each time step, and we have a couple of methods for doing this. One method is to use online expectation maximization, which is a current common algorithm for holistic models. Another algorithm we use is gradient descent using the gradient of the log density of the observations which is computed as a byproduct of the bird.

F

The message passing process, but probably the most interesting kind of learning, is what we call a productive hypothesis generation and what this says is that suppose we encounter evidence that doesn't fit any hypothesis that we currently have like the hopping vehicle. So we have to generate new hypotheses that reconcile experimental data from the layer below, with theory from the layer above and we're developing algorithms, to do this and we're trying to admit a scientific hypothesis, generation and validation process.

F

Where, given the current hypotheses, we hypothesize and new alternative hypotheses that are perhaps a small Delta away from our current hypotheses. And then we see if any of those matches the data and and so we could potentially search for a lot of different hypotheses. But we have a preference for hypotheses that are nearer to our existing hypotheses.

F

D

Abby there's a question from the audience: how does the system generate hypotheses.

F

Yeah so right now it's completely ad-hoc, so associated with each node. That is doing this hypothesis generation. We have a hard-coded hypothesis generator that basically is a function that takes one hypothesis and returns another set of hypotheses with the probability distribution, because this is very early work along these lines. But what we're aiming to do is identify standard ways of.

F

Proposing hypotheses about something so, for example, let's say you're proposing a hypothesis about a numerical quantity. Well, one standard thing you might do is to propose that, okay, that quantity changes over time according to some inflation function, or something like that. So if we can discover standard things like that, we can learn how to propose hypotheses, but it's interesting because maybe you could actually have the higher levels of the hierarchy indicate how to propose hypotheses.

F

So there's a lot of thinking to do here and a lot of what to do here, we're just at the very early stages of this hypothesis generation. So, where we've applied it we've had a challenge problem which is to play Monopoly.

F

When someone is changing the rules on you and not telling you about it, so suddenly they could change the distribution on the dice or they could change the number of locations on the board or they could change the prices of properties and so we've hand coded hypothesis generators for all those different kinds of changes, but we'd love to find a systematic way to do that.

D

Okay, great thanks, I think one more question and then I'm going to let you go all the way to the end. I think we have about. We should leave about five more minutes for your talk and then the rest of the questions, but the one other questions we have from the audience. How do you construct analogies amongst the domain-specific generators.

F

So that's a great question and I: don't really know the answer to that yet, but.

F

There are certain questions that that get asked a lot like how many things are there right and you, sir, and you originally think that that number of things lies within some distribution and then you see something that's outside the distribution. So if you have a change to a how many things question, there could be a general way of proposing a new hypothesis like let's say you think, there's okay, there's between one five things and then you've encounter seven things now. Are you gonna propose that, oh, that could be a hundred of them?

F

Probably not you're gonna propose that there's some number maybe up to ten something like that right. So are there ways that an AI system can do that in a way that leads to good results and there and the same thing about other kinds of questions as well like changing distribution. So are there other kinds of distributions that would be natural to propose and when the distribution changes things like that, yeah.

D

That's one of my questions for you at near the end as well. So let let you finish now we'll take the rest of the questions during the discussion period. So if you could take about five more minutes and then we can switch to general discussion, that'd be great I'm.

F

Sure I only have a few more.

D

Minutes so I'm.

F

Gonna skip this. This just shows some kinds of ways we can do hierarchical learning, but I've talked about this already. I want to spend a few minutes talking about representation, so we need representations of layers of the hierarchy that are expressive enough to represent beautiful models and flexible enough to learn interesting functions from data going back to my original point of not losing the ability of deep learning to really learn, complex, rich functions that describe the world and they also need to be accurately inferable with belief propagation. So we have two methods that we use.

F

One is to use representations that are provably tractable, while maintaining layers of latent variables and another one is to learn, representations that pretty well under asynchronous belief.

F

Propagation I'm gonna mention a couple of ideas for the first approach, so in these hierarchical representations we have layer functions which are vectors of noes, providing a distributed encoding like a neural net and individual node conditional probabilities that use for ballistic programming, and we have several example- representations that we're developing, such as separable models, deep, noisy or networks, deep conditionally, near Gaussian networks and for actual node tree with models and I'll just mention a couple of these.

F

So separable models are models that the internal nodes satisfy this mathematical equation, which I'm not going to go into, but it's just a particular structure of the node ordinarily, when a variable X depends on parents. You warm through UN, there's no way to decompose this representation, a separable model decomposes it additively into a weighted sum of individual dependencies on each of the parents, and we've shown that you can get very fast and accurate inference with belief propagation.

F

In this way, and so as a result, this gives us a method for building deep, deep networks with separable models that can be in use that you can use belief, propagation and we're currently running experiments to see how well these can actually capture different kinds of data and starting to get interesting results.

F

The other kind of model I want to mention is the deep, noisy or network, so noise.

F

Your models are classic models used in Bayes nets for modeling courses that act on their effect independently, so we have multiple causes and each cause leads to the effect using an or bottle of disjunction, except there's a noise term associated with each cause that might prevent the calls from achieving the effect, which is why it's called a noisy war and a deep, noisy all Network is a layered network in which each node is noisy, or so there are a couple of interesting things about deep, noisy or networks.

F

First of all, if you observe a false effect, you can infer the causes independently, which means it's very efficient: to perform exact inference. True throughout the network, using a single BAC would pass, whereas if you observe a true effect, all the causes become coupled leading to exponential inference, on the other hand, in general for a sufficiently wide and deep, deep, noisy or Network. The prior probability of true for all these will be very high. So a true observation is not prizing provides little information.

F

A false observation is very surprising, provides lots of information which suggests an approximate inference. Algorithm ignore the true observations and process the false observations in linear time.

F

So we can think of a deep noise. You will network as an anomaly detector, or you might actually want to use a noisy end, because it's the inverse of noisy or- and you can determine what's true so suppose, you've observed two anomalous observations, so these would be true and everything else would be false. You only process those in linear time to determine what is true at the root, and you ignore the false observations, and so by determining that this thing with the exclamation point of the root is true.

F

You've detected an anomaly and you've actually said what that anomaly is, and this is interesting because it has a predictive processing interpretation we're in an anomaly detection situation. Our prediction is always going to be that there's no anomaly, we're always going to predict false for these anomaly nodes and an error is when we observe that something is true, and then we propagate those errors backwards in linear time to determine the causes of those errors.

F

Okay, so conclusion why this might help for human-level AI, we can build the right abstractions. It's both knowledge, driven and deeply learn about continually learns and improves, and the act algorithms we use a fast, so we're not claimed to have exact inference which is really intractable, but these are approximate but realistic. We hope, algorithms and I think I'll end there to allow some time for questions. Okay,.

D

Great thank you so I'm going to unmute Jarek as well. Let's see if that works.

D

Okay, I think we had at least one question from the audience: I'm going to unmute Ekrem piece go ahead and ask a question. Sure.

G

Yes, ah thank you. First of all, thank you for both of you for this great presentations. Actually, my question is more like a commentary on how do you think about this, this prediction frameworks fit into the picture of active inference or free energy principle.

G

Professor Kristen's work or line of work. I would like to hear both of you. What do you think about it?

G

D

Okay, who wants to go first.

C

Happy to give it a go, um so the first part I'm, not sure, I, understand active inference in what sense, I'm, not sure I, understand the framework. I mean Bayesian, inference or yeah.

G

It's actually built upon the Bayesian inference, but it is more depend on the free energy principle on yeah. That's I would like to see how does theme fit in the picture that form vice versa and active.

D

Inference sort of specifically includes kind of the sensorimotor component, the motor component, if you take your actions into account as part of your kind at the overall predictive model, meaning I.

C

Mean in general, I guess: I, don't know enough about active inference, but in general the idea of Bayesian inference would be that you have a application of probability. Distributions I mean, as the principle says, and so the difference to this is I. Think of predictive processing is a biologically plausible implementation of the more mathematically, accurate and beautiful version of of Bayesian inference.

C

So that would be the relation to active inference and then I mean to be completely honest about the free energy principle. I just don't know enough I'm, not sure I understand it fully. I can comment on at least the assumption that predictive processing is more efficient in terms of spikes and and that's an assumption.

C

One reads about relatively frequently a surprising one, because I don't think it's true I mean it's it's true in a very, very specific subset of networks, and the simple reason is that you can't signal negative numbers in neurons, so you for every every sign inversion that you invariably need for predictive processing. You need an extra neuron that fires an extra spike, so every top-down spike needs to be canceled by a bottom-up spike or vice versa.

C

So it's not it's not more efficient in number of spikes, so I think a lot of the at least you know some of the assumptions that go into this might not be it's completely accurate.

D

Okay, thanks AVI: do you have any anything to add.

F

Or compatible with friskin's thinking, but maybe don't go as far as he does in sort of unifying everything under this free energy principle in particular, one thing I neglected to talk about because we haven't done a lot of work on it, but I think is very important, is integrating action into our framework. For the most part, we've been focusing on perception and reasoning, and for action we integrate with a planning model or a reinforcement, learning model, one of the interesting ideas.

F

I think this is maybe the active inference idea, which is but the idea that in predictive processing you act in order to reduce your surprise. So your goal is to get into a state of the world that you're not surprised at all, and you take actions to reduce your surprise and that's an interesting point of view. That would be interesting to think about integrating into our approach.

D

Okay, yeah kind of the standard paradox there is, you know if you're, mmm if your primary purpose is to kind of act to reduce your surprise, you know why don't you just go off into a corner turn off all the lights and just say that yeah great.

D

So just a reminder, you know if you want to ask questions, just please note so on chat and I cannot meet you or I. Think you can raise your hand if you go to the participants tab and there should be a button there to quote-unquote, raise your hand as well RS or Lucas I. Don't know if you had any other questions to ask and I thought I'd have a question I think.

A

Arthur is hazing with hand Oh.

D

Arthur's singer: yes,.

H

Thank thank you yeah. This converges this converges with what we've been discussing, but coming from a different level. I've been studying deep learning, and it's striking that within the last couple of weeks, they've just announced the complete dissection of the mouse, prego million neurons or so and I. Look at that and look at the sudden surge of great success and superhuman, deep learning and look at the two and say these two are converging and I think that applies even more so to the sort of stuff we've been discussing here today.

D

Yeah, let's hope so that's definitely there. It's did you guys want to comment at all.

C

That, yes, no it's fascinating, it's just maybe I-43 accuracy, I, don't think it was completely. The conic thumb of the mouse is not solved yet I. Think they've made some progress in in going towards their cubic millimeter, but not not. The entire nervous system, which would be also is considerably more than a million neurons yeah, even even in the mouse.

D

Yes, I have a basic question kind of on the you know the nature of the hierarchy. If you will so Jeremy, you kind of ended your presentation on a pretty provocative note that you know, through your work on predictive processing, you think the hierarchy may not even be necessary, at least in the visual cortex may not even be present and then avi in your work. You know it was very hierarchical.

D

You know the hierarchy is critical to work and I. Just wonder how you, both of you, think about kind of the nature of the hierarchy and predictive processing.

C

Right I mean maybe to to clarify I. Think it's not from the argument is from basic principles, not from a lot of data that the whole the system as a whole likely is not already hurricane. um This doesn't mean that locally, it's not very useful to have a hierarchy, and even if you look at all these diagrams, they they're- you know, there's these four parallel hierarchies and they they communicate right. So how do you? My point is that you can't put the whole thing into, or you know that you you it's computationally, probably not advantageous.

C

To put the whole thing into one global hierarchy and cortex I would argue. Certainly isn't you can certainly find evidence for local hierarchies, as we have been doing, we being neuroscience for four decades. Looking at you know, processing higher is in sensory streams and cortex, and that certainly has its merit, but the system as a whole likely is not the ranged in such a way that it can only function as a hierarchy. As the very simple you know, toy and model of two brain areas.

C

One sensory one motor already demonstrates it's very hard to define reasonable hierarchy and such as.

D

Yes, yeah: go ahead, Abby I think.

F

Having everything in the hierarchy is not so important, but there are hierarchical relationships that I think are important, so general to specific abstract to concrete halt apart. All those I think are fundamental kinds of hierarchical relationships, also I'm speaking as an engineer, not a neuroscientist. So if I think about how to design a system that can reason and learn, I'm gonna think about those kinds of relationships, and that's really what what's important to me in the hierarchy: I, don't care that everything is in a in a whole line or anything like that. Right.

C

I mean I think the other advantage of having not a strict hierarchies is much easier to integrate new regions. You can build, you can build your system, module early I mean I. If I'm not mistaken this. This is sort of the idea that Jeff has where there is a thousand brain, all right, one of the ideas and probably missmiss representing the idea, but that you can modularly expand that you don't need to rely on a hierarchy where you need to build your computational levels into a pre-existing re. A computational models, sign module.

E

I think that's a Jeff speaking yeah! That's right! You know. The cortex looks like it's locally got some hierarchical components and you can see the anatomically there's different types of connections between regions. It's some fall into those sort of hierarchical scheme, but the majority of connections do not at all, and so one of the things that are the thousand brains that posits it's at least an understanding of how the disparate regions, who really don't know anything about each other, can communicate through voting.

D

Yeah most struck even going back to set of fellowmen and venison's paper. Even in that paper they noted that you know, although they they showed the diag recordable regions in the hierarchy. They note that 40% of all the possible connections that could exist actually do exist. So, instead of going back to what you said, the arrogance, it's you know you could find a hierarchy if you look for it, but it's not necessarily it's.

E

That, even worse than that paper came out, the original paper came out in 91 and they didn't even look for connections like between a1 and b1 or s1 and DUI. It's just. They didn't even think that those were possible. But now we know those connections exist and they really.

D

E

Really defy the whole concept. Well, those connections find accounts of the parties. I think George is corrected, that it's a it's a bit above, but I think through think about the hierarchy in its classic form. It's a mistake.

F

Well, one thing I want to add is that having the hierarchy does not preclude the possibility of feedback edges that go from lower levels, to higher levels, right, right and, and so that might be a way to say that we have a causal flow that tends to go a certain way. But there can be these feedback edges that inference the higher levels and make everything as if it's one.

C

Yes, don't know, of course, as soon as you relax the strict adherence the higher. Even it's just a question of how much you relax it I.

E

Think a big pool again is this anatomically. There are different, physically, indiscernible different types of connections between regions, there's ones that go through the thalamus and there's ones that go directly and they originate different layers. And so this can tell you anatomically the types of relationship regions can have, and only some of those.

D

There's a pretty good deep topic and you know one of the questions that we've studied a lot of are kind of the lateral questions you know within a region or within the same level, quote-unquote level. There are, you know, lots of sideways connections and I, don't know avi if you have any sort of an explanation for kind of lateral, quad connections in your you know.

D

F

F

Fundamentally, our approach is based on directed graphs like Bayesian networks, so all probabilistic in influences are from parents to children. However, there may be, because this is a dynamic model, there may be an influence from the child to the parent of the next time, step and they're, essentially at the same level. So even even though, from a mathematical point of view and the way we can make this efficient and the way we can make this work, everything is very active.

F

It doesn't mean that we don't have many variables that are influencing each other all the time over time and are effectively at the same level.

D

Stephane, did you have a question? Yes,.

I

I was curious. It is the exact definition of the hierarchy which is used to discuss it, because I think hierarchy means different things to different people. So to make sure that we talk about the same thing and precise definition would be useful right.

C

So I mean you're, absolutely right, intuitively it's easier to understand that there's just some absolute level level of abstraction that increases at least in neuroscience when one thinks of a hierarchy, but in principle you can think of it as a relationship. You can define between any two areas where the relationship is either equal, lower or higher, and you never run into contradictions for the entire network. I think, then you, if you can set that up in such a way, there's no contradictions. Then you have a global hierarchy that is consistent. Okay,.

I

So we did, we build a mathematical model based on the measurement of the Manzo's copy connectivity in the mouse brain, using a definition very similar to that. Just the question of doing an unsupervised. Clustering of connections such that they're, both feed-forward and feedback connections, and try to see how self-consistent are the relations. If we define particular types of connections, feed-forward particular types feedback- and we found out that, yes, we can define hierarchy, but the level of consistency is super. Low.

I

I would say that only a small that there is a small bias in the total number of connections which are preponderance in one directions. There are many connections which don't follow this so just based on a non analogy. We computed. So imagining that perfect consistency between two floors would be a level of 1.

I

We found that 2 is about many 60 areas in talamo's, plus cortex. We found the equivalent of 2 full levels, so the hierarchy score is about statistically average, but it is way shallower than what would one would expect by putting all the areas one on top of each other. So it is I would say that, yes, we have a hierarchy based on that mathematical score, but it is the vast majority of the connections actually don't end up following it. There is just a small statistical probability on top that paper came out last year in nature.

I

I could put a link if needed, I.

C

Saw it very nice work yeah, um no I I would fully concur with this I mean my. My original rant was more against the idea of a strict hierarchy. Where were they where everything is organized as a hierarchy, which is an impression one can get? If you know from certain neuroscience.

H

C

Older neuroscientists work, neuroscience works, yeah.

D

And also from deep learning, you know which.

C

Has been exactly which has been one of the core ideas that has proven to be very successful and I mean if this isn't it isn't to say that there isn't the very important use of having hierarchical relationships. But I think if we want to progress beyond this state or at one of one of the options that is probably worth pursuing, is the idea of networks that are not hierarchical or at least partially non hierarchical.

E

Stefan the worked retirement was that math mice or rats. What was that it.

I

E

So I mean yeah, you might you it is it possible you might get more of a hierarchical representation in a primary, for example, I mean, obviously you know in a mouse. Almost all vision occurs in one region, maybe one its large and any subsequent hierarchical teeny. So I just noted, if you think that result would apply to more complex brains, I.

I

Do think we we sent the code which we used to mathematically compute. So do you think that it was one thing which was useful there to actually have a mathematical definition of hierarchy and we sent a code to some people who do crime at research, but I haven't heard anything back to try to see. I would be so curious to see how how hierarchical it is in other animals, but we do see a difference.

I

So there are I think about nine visual areas in the mouse, but we don't the they are hierarchically organized. But just as Burke said before the higher the many connections don't follow. The strict rules of the hierarchy. You've got a basic like a backbone of hierarchy and then a huge sluice of connections which don't follow. It am.

E

I right that in the mouse, v1 is relatively quite large compared to your cortex and all the subsequent visual readers are very small. Is that still correct? Yes,.

I

That is correct. The entire volume of visual cortex I think that almost a half is occupied by v1, not going not almost yeah.

E

Since it what I take from that in the hierarchy is not very deep and so I taking, that is that most of vision occurs in the mouse and B. What I mean you should look at the amount of material there and see that's one of the things that we kind of drive. Our thinking is like say: well, it's so much is occurring in this primary visual region. It can't be just doing something quite simple: it has to be doing a fairly sophisticated model.

E

At least it's it's as much as I'm asking models of world vision fully.

I

B

I have a question for your. If, if nobody else has a question, go ahead, so you're well, I kind of have one and a half questions. One is what is the exact separation functionally between feedback projections, or you know, project, predictive projections between cortical areas and the predictive, the putative predictive capacity of deep layer neurons and also what are deep layer neurons that target the thalamus or the or motor regions like the superior colliculus? What do you think those are doing in this context?

B

C

An interesting question I mean as I try to do too in the talk. We have this idea, that info granular layers or deep layers and cortex function to maintain this representation, but we know very little about it and there's many different ways. This could be implemented in principle.

C

Now the question becomes, you know what the oh sorry! In addition, what you need to having these representations, these coordinate transforms. So you need to send the information between the areas and and then but this needs to be transformed. Somehow, and usually the assumption is that the weight matrix of the long-range connection is what does the transform and and there we have a bit of data on the interaction between a1 and d1, where it looks like.

C

So, if you, the experiment were doing, is recording, but this is the Lena Garner postdoc and lab doing this experiment and what she does is she records from the a1 axons in visual cortex during paired audio-visual presentations, and it turns out that the response is in the a1 axons. Look like they're visual signals. They don't depend on visual cortex, so she can silence visual cortex locally, and this is very little effect on these visual signals.

C

So this has has fundamentally changed the way, I think about this interaction and how the how the predictions are sent between the areas, because it looks like the signals, are transformed before.

H

C

Sent to the target area so the transform, at least in part, it's not that there aren't any auditory signals in the accident, largely visual ed, and it looks like the transformation happens before they're sent out. So a lot of the circuitry likely is devoted to not just maintaining the internal representation, but also transforming the signal before at the cortical area, sends it to its target area the predictions and and that all of that needs to be also integrated into.

C

You know, layer, four, five and six, assuming assuming that 2/3 actually functions mainly to do the prediction error computation, which is a obviously.

H

C

I'm not sure that answers the question is or the one in.

B

Our that was great thank you.

D

Okay, it looks like Stefan put the link to the paper in the chat, so you I should go grab that any other questions from anyone. I recognize we're running a little bit late beyond any other questions from anyone in the audience.

H

G

B

But I have a question for obvious as well. How do you mentioned gehrig's mention of learning forward in inverse models in in scrubs in your probabilistic framework? How? How do you think you could learn inverse models like this? The richness in representation provided by a probabilistic programs help here because it's a you know it's a very hard problem. I was wondering if you have any insight on that. Yes,.

F

Good question so one of the main motivations for probabilistic programming and main applications has actually been inverse models so, for example, inverting a graphics engine if you have a probabilistic model of a scene and then a graphics engine that renders that to an image inverting that to recover the scene from the image. That's not learning the inverse model, though, but if you have a model- and let's say it's a rich model- a forward model to take the example, a model of how vocal actions cause certain sounds to take the example from gigs tour right.

F

So how do vocal actions call certain sounds it's gonna be a rich model which has a certain amount of code to work, trouble asleep programming code, and when you get observations you can compute the derivative of the probability of those observations with respect to parameters of that code and then by adjusting those in the direction of the gradient. You can learn a model that would have made that observation that sensation more likely of one of the ways Bayesian update, is another way of doing that right.

F

So if you have a hypothesis at the high level of how your motor actions will generate certain sounds using Bayesian update, you could in you could increase the probability of those hypotheses that make those sounds more likely. So in both those ways, your training, the inverse model, the forward velocity you're, not training in this model, you're trading, the forward model.

B

Okay and so but but that wouldn't help you so basically you saying that there's an inverse model, basically essentially by definition, since you have a forward model and you can reason backwards so.

G

F

G

Really learning.

B

F

That's why you're not learning an inversed, nor when you're learning, okay, you're learning the forward model now some people for the purpose of making inference efficient will train an inverse model from the from the forward model like in your network, or something to do efficient, inverse inference. That's just an inference technique. It's not really learning, even though the learning algorithms will use. It's not really learning anything new about the world that you didn't know before. So the learning itself is learning the forward model.

B

D

Yeah great, thank you and there's one more question from Dean Ron's. He says he can't use the mic right now. This is for Georg. So how would a attention fit into the predictive coding framework if it changes neuron outputs? Wouldn't that mean it changes prediction errors? How do you think about attention in your in your overall framework? Yes,.

C

That's an excellent question: I mean there's a there's. A two answers to this. One is a simple one. You can obviously use the prediction itself to to gate. Well, you know, or to do something like attention to change the strength of the predictions. The other thing that we think cortex might be using is using their module. Ettore signals to to gate how sensitive layer of an internal representation neuron is to the prediction errors and it bit like the data showed with the antipsychotics I. Think I think you can modulate this fire direction.

C

I mean one of the signals were looking at this cholinergic input. I mean you start working model right now, I mean this is all the grain of salt. Is that that the cholinergic inputs change is the way layer? 5 is influenced by Larry two three, so you all coagulate that and then layer 5 becomes more or less sensitive to one or the other type of Larry to three input: positive or negative prediction errors, but that would be.

C

In theory, I mean just to give a very speculative answer: interior, where you could integrate things like more global, attentive signals right if you want to down-regulate visual input and I'll percolate auditory input without changing the reps. Without you know, disturbing the internal representation completely, but just gating how it's updated, whether it's updated more from the periphery or more by the internal drive or top Dondre. We.

D

Could cholinergic modulators work at a fast enough time scale tension related up? It affects.

C

Probably I mean we've recorded extensive, we can call them neurotic signals and visual cortex and there or on the same scale as the input I mean that they're, relatively quick to onset I mean remind you. This is our population signal. So it's it's right there it's hard to say, but there's definitely you know on the sub-second timescale.

C

D

Be do you think about attention at all any comments on that.

F

It's important but haven't got there yet that there's there's so much to work out and I know. Attention is important.

F

G

D

Fair enough, okay, well we're running very late here and I really appreciate everyone's time. I know for for the speakers it's evening. Your time and I really appreciate you guys being around. These were excellent talks and they were recorded, and so we will put a version of that. It'll probably take us a few days.

D

Matt, who was supposed to be the moderator I had to jump in last minute, couldn't make it unfortunately, but we'll try to get it up on our on our YouTube channel at some point, so thank you very much again and thanks Garriga nabi. We really appreciated you joining in today. Thank.

F

You very much it was a pleasure all.

D

Right yeah take care, bye, bye, bye,.

D