video_id
stringlengths
11
11
text
stringlengths
361
490
start_second
int64
0
11.3k
end_second
int64
18
11.3k
url
stringlengths
48
52
title
stringlengths
0
100
thumbnail
stringlengths
0
52
P0yVuoATjzs
does these days but I recently also took a gig as the director of a new MIT IBM collaboration a quarter-million-dollar AI Institute so if you if you're interested in that come and find me and then I'm going to tell you about some deep learning work we've done but then I'm also going to tell you at the end if I have time about some experimental work that it's inspired in
36
54
https://www.youtube.com/watch?v=P0yVuoATjzs&t=36s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
my lab just to kind of reinforce this idea that there's sort of a loop that we can be driving between models and experiments so we all know that deep learning has kind of a maybe kind of connection to biology so we have units and they have synapses and connections between them okay that's that's part of it that's for their artificial neural network part comes from and then they're
54
75
https://www.youtube.com/watch?v=P0yVuoATjzs&t=54s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
deep and you know we know that perceptual hierarchies for instance in the and the primate are are sort of deep hierarchical systems and deep CN n sort of capture that but really when you get right down to it you know that's been most of the interplay between these two and I think a lot of us here are sort of trying to think well how can we go back to the brain and get more inspiration
75
95
https://www.youtube.com/watch?v=P0yVuoATjzs&t=75s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
but then also how can we use the in you know the deep learning to actually help us understand the brain a little bit better so again this virtuous loop you're gonna hear this I think again and again and again and that's that's wherever we're coming from so it turns out the deep learning so I'm interested in perception so a different part of the stack then the the previous two talks
95
112
https://www.youtube.com/watch?v=P0yVuoATjzs&t=95s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
imaged in how we perceive objects and make sense of this sort of terrible complicated flow of information coming in through our senses and it turns out that by accident deep learning systems convolution neural networks turned out to be the best model for doing this so people have built all kinds of computational models of how visual processing worked in the ventral visual
112
132
https://www.youtube.com/watch?v=P0yVuoATjzs&t=112s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
pathway and primates but it turns out that when you just started training deep Nets you just took the internal representations compare them to actual neuronal population responses and that's the best model so so here this is a paper from my my former PhD advisor Jim - Carlos lab and basically you see those are sort of model fit quality for different kinds of models that came
132
151
https://www.youtube.com/watch?v=P0yVuoATjzs&t=132s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
before including some of my own and then basically once we have deep Nets those explained the data fit the data way better than anything else and you know this has been going on and basically the bottom line is the better the models get on image net the better the fit seems to be between what's going what the representational space looks like in the in the deep net and what the
151
169
https://www.youtube.com/watch?v=P0yVuoATjzs&t=151s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
representational space seems to look like in the population so this led to this idea that well maybe maybe visions just tapped out and this is a solved problem and then maybe we don't need to worry about this anymore the answer is deep Nets there's a problem because all the work we've done so far is all looking at static representations and really our visual systems are built for
169
186
https://www.youtube.com/watch?v=P0yVuoATjzs&t=169s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
dynamic situations and it's even worse than that because even if you show a static stimulus neuronal populations will not produce static outputs and this is one of the first things you notice when you're a graduate students sticking electrodes into the brain of the monkey which I did for five years if you show a static picture of a monkey face you get these sort of transient
186
204
https://www.youtube.com/watch?v=P0yVuoATjzs&t=186s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
responses and in fact the only way to drive a sustained response in an IT neuron at the end of the ventral visual pathway is to show a dynamic stimulus so this is very much at odds with how cnn's work because they have no intrinsic notion of time you put in an input you get an output it's a static thing so something is clearly going on here and there's all kinds of interesting rich
204
220
https://www.youtube.com/watch?v=P0yVuoATjzs&t=204s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
dynamics here where sometimes it's sustained sometimes it's not sometimes it looks like it might be oscillating so we don't really have a good picture yet of what that's all about and to that and you know that part which seems really salient isn't actually captured by simple cnn's and it gets even weirder than that too because there's this great experiment by Karl Olsen's lab where he
220
239
https://www.youtube.com/watch?v=P0yVuoATjzs&t=220s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
showed successive presentations of image image image image and in some cases the the second image the B image was predicted by the first image so b4 always happened after a4 and you saw this a couple hundred times and b5 was always seen after a5 showed a couple hundred times and then basically you see is you don't get a response to beef-up before if it's preceded by a for does
239
260
https://www.youtube.com/watch?v=P0yVuoATjzs&t=239s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
that mean that the cell just doesn't like b4 well no actually it will respond to b4 but only when it's preceded by a different image so there's some kind of higher-order temporal track that's going on but again CNN's just don't capture because they don't have any intrinsic notion of time the other thing too which i think is actually a salient problem for deep learning is
260
279
https://www.youtube.com/watch?v=P0yVuoATjzs&t=260s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
that most of the datasets we have to train these things you need tons and tons and tons of data you know if you want to train you know a dog detector you need you know thousands and thousands and thousands of dogs and thousands and thousands of thousands of things that aren't dogs and that's just not the way we learned I don't sit down for my daughter and shower dog dog dog
279
295
https://www.youtube.com/watch?v=P0yVuoATjzs&t=279s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
dog dog Cat Cat Cat Cat that's just not the way it works in fact yeah we can just do that experiment right now does anyone know what this is raise your hands if you do okay we have a few electrical engineers in the audience okay even though you've even though you've only seen this for the first time is it present in this image yes how many in this image - how about that one yeah
295
315
https://www.youtube.com/watch?v=P0yVuoATjzs&t=295s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
but it's a little weird right so even though you only saw one example you were immediately able to determine what was there you immediately an expert on this kind of object that's called one-shot learning and you know just if you need more evidence that deep nets don't quite work yet is anyone can anyone say what this is it's actually Meret Oppenheim it's called luncheon in fur but everyone
315
338
https://www.youtube.com/watch?v=P0yVuoATjzs&t=315s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
agrees it's a cup a saucer in the spoon state of the art deep net says it's a teddy bear and then it gets worse than that even even when things are very clear to the right object so this is a state of the art our CNN detection correctly detects this as a bird but you just put a few objects in the don't belong and it starts you know calls this a cat says that's it you know it's this
338
359
https://www.youtube.com/watch?v=P0yVuoATjzs&t=338s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
is a television but so it's calling that a cat so there's something about there's something brittle and adversarial examples we could go on and on and on about ways in which deep nets really aren't quite solving the problem yet and you know the I I and other people think that unsupervised learning is an important piece of this so how can we without labels gather up a lot of
359
378
https://www.youtube.com/watch?v=P0yVuoATjzs&t=359s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
structure about the world and build representations that are really good and then that can feed into things like reinforcement learning and all that kind of stuff so this is just a different part of the stack I think we all agree on there's multiple kinds of learning happening here but the kind that I'm really interested in and have been for a long time is this idea of temporal
378
394
https://www.youtube.com/watch?v=P0yVuoATjzs&t=378s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
learning if you just look at the environment and let it play out the environment is almost always showing you its structure so you look at a person doing a tennis serve you can see how they're articulated you can see they're put together you can see how the shadows move around on an object that's just all played out in time you don't need supervision you can just
394
409
https://www.youtube.com/watch?v=P0yVuoATjzs&t=394s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
observe and learn a lot about the world so that's what we're interested doing it and you know brains from the from the neuroscience literature seemed to be exquisitely tuned for these kinds of temporal statistics so this is one of my favorite studies it's like a physics where basically they took faces and then they rotated them and then had subjects to passively watch these but as a little
409
428
https://www.youtube.com/watch?v=P0yVuoATjzs&t=409s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
trick sometimes they would morph the face as it moved and people generally didn't notice this but if you ask the same different later ones that had morphed and they had seen them morph a few times they would incorrectly sort of associate them as being the same same object and then you can even go further this is let me I didn't over my PhD you can even have presented a peripheral
428
446
https://www.youtube.com/watch?v=P0yVuoATjzs&t=428s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
object and then have while the subject is a Codding you flip it out for a different object before their eyes land their eyes land on a different object and then you ask them to do same different on peripheral versus fovea and they'll make incorrect associations so it seems like the brain is constantly collecting up these sort of temporal associations to make sense of the world
446
463
https://www.youtube.com/watch?v=P0yVuoATjzs&t=446s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
so we have we have experimental of us this might be happening but we don't have really models for it now there have been prior attempts so slow Fourier analysis is a is an attempt to sort of extract those signals that are moving slowly and this has been influential but you know it hasn't sort of unleashed you know a revolution in how we do things but it's an interesting and important
463
482
https://www.youtube.com/watch?v=P0yVuoATjzs&t=463s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
set of ideas but what I want to do is look specifically at prediction as an unsupervised learning rule so basically the idea that can we just use the idea of predicting what a future frame is going to look like to help us build better representations and I'm gonna make the argument that brains are particularly adept at prediction I'm gonna use this person who's this person
482
501
https://www.youtube.com/watch?v=P0yVuoATjzs&t=482s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
Sarina yes not Venus Sarina you guys are our experts so anyway so she can she can do tennis serve at 207 km/h that's 57 meters per second that means the ball traverses the court in 400 milliseconds about now the latency from the retina to the primary visual cortex is about 60 milliseconds so that means that v1 is operating 3 meters in the past and if we go all the way through the ventral
501
524
https://www.youtube.com/watch?v=P0yVuoATjzs&t=501s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
visual hierarchy then we're talking about about a hundred and seventy millisecond latency in a human it's about nine meters so there's a very deep sense in which if you think you saw Serena Williams's tennis serve you couldn't have because your brain you know this year line latency of your brain was way in the past so one suggestion that that's been made is if you're returning that
524
543
https://www.youtube.com/watch?v=P0yVuoATjzs&t=524s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
serve a big part of what you're doing is you're looking at the lined up and you're predicting where the ball is going to be and that's where you're putting your racket so you know and this this idea has been sort of you know it can be shown there's a little bit of nuance here but if this is called the flash lag you lose inch so if you look at this dot right here and then watch
543
561
https://www.youtube.com/watch?v=P0yVuoATjzs&t=543s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
this sort of clock hand going around you see how there's like a little flashing straight line now is that line behind in front of or lined up with the clock hand behind of course and the rule with all optical illusions is if it looks like it's not lined up it is lined up and if it looks like it is lined up it's not lined up so so this is perfectly collinear and the sort of classical interpretation
561
583
https://www.youtube.com/watch?v=P0yVuoATjzs&t=561s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
and there's some nuance here is that basically this line has to go through the full pipeline latency of the visual system but this tracking line is predictable so your visual your perception just sort of puts it where it really is in real time so so this is and then there's another weird things everyone's kind of see that it looks like it's tilted as well I can just hold
583
601
https://www.youtube.com/watch?v=P0yVuoATjzs&t=583s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
that thought all right so so so so we wanted to get into this so we were we were you know working in sort of machine learning computer vision and right around the time we were working on this we got interesting this idea of future framed video prediction so basically the idea is that if we see a sequence of images transforming the goal is to basically predict what the next image
601
623
https://www.youtube.com/watch?v=P0yVuoATjzs&t=601s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
looks like and this is now that there were there was a little bit of work around the time we were doing this and that's it's a much bigger field now just to acknowledge that there are other people working in this area and we did some very simple things that you would do in in deep learning and again this is sort of fits in the same spirit of what Matt was talking about where so you just
623
639
https://www.youtube.com/watch?v=P0yVuoATjzs&t=623s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
want the simplest possible models to start with and just see how much your problem statement sort of you know sort of it makes the problem happen so basically we did is we took an autoencoder type architecture and we just wedged a recurrent neural network in this case an LS TM in the middle and then asked and then we also had a fancier version with gans back then
639
657
https://www.youtube.com/watch?v=P0yVuoATjzs&t=639s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
there were only four Gann papers so it seemed exotic and how there's like 4,000 Gann papers but anyway so basically what we discovered is we can build these sort of generative networks that could actually basically sort of renderer faces and this doesn't seem so surprising anymore but at the time I was shocked how well this worked and there's some details about whether you
657
674
https://www.youtube.com/watch?v=P0yVuoATjzs&t=657s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
use the Gann or not and whether the ear appears or not but the basic bottom line is this is there a future frame video prediction problem where we see a sequence of frames and then we predict what the next frame might look like it is actually a tractable thing we can do with neural networks now and then the interesting and of course it works on all kinds of different faces including
674
691
https://www.youtube.com/watch?v=P0yVuoATjzs&t=674s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
faces you haven't seen before but the interesting thing here was we were really trying to look at what kind of representations do we implicitly induce when we train a network to do this kind of future frame prediction so all we're training we're training the network with back prop but the loss is coming entirely from how well does it reconstruct a future frame that it
691
707
https://www.youtube.com/watch?v=P0yVuoATjzs&t=691s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
hasn't seen before and the interesting interesting thing that we find is if we kind of just take that internal representation and just sort of pipe it off and just ask how well can we decode things about other things like the identity of the face what we find is if we do a simple you know 50 Way face recognition task where we need to tell which of 50 people it is and we only get
707
728
https://www.youtube.com/watch?v=P0yVuoATjzs&t=707s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
to see 1 2 3 4 so on views of the person what we see is if we start using these predictive networks they're able to do a little bit better with a little bit less data and this is what we're trying to get towards can we get representations where we can do more with less training data and because we're sort of extracting some sort of sort of deep part of you know some sort of deep
728
748
https://www.youtube.com/watch?v=P0yVuoATjzs&t=728s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
structure of the image now this was back in in 2015 and then you know we also discovered that you know for instance you could internal within to these the representations that we learned you could also find directions that you can move around in that would do things like make the face more male or more female these are less surprising now that ganzar around and we can do all kinds of
748
769
https://www.youtube.com/watch?v=P0yVuoATjzs&t=748s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
fancy things with them but what we really we're driving towards just to get more to the punchline was not this sort of very simple auto encoder with an RNN in the middle but really getting towards this idea of predictive coding which came from neuroscience so the basic idea here this is a paper from 1999 that popularized predictive coding but the idea actually goes back about a decade
769
788
https://www.youtube.com/watch?v=P0yVuoATjzs&t=769s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
earlier and the basic idea is this we start with an input we have a feed-forward signal we try and predict away the input and we subtract that off and we only send forward the differences now the original idea here was an efficient coding idea let's try and reduce the number of spikes we have to send because if we subtract away the things we already know then all we need to send forward is the
788
810
https://www.youtube.com/watch?v=P0yVuoATjzs&t=788s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
is the the difference from that so what we did was basically to take yeah on what we think the signal is if I'm looking at the outside and I'm trying to submit position of a ball and predicting where the ball is going to be is a different stories and if I'm trying to estimate suppose it's octave all right and then do in how many examples do we actually know what it is that we're
810
835
https://www.youtube.com/watch?v=P0yVuoATjzs&t=810s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
trying to predict I mean really for sure right this is only doing that thing right yeah that's that's a great matter question and and I think we can discuss that like what are we estimating in this case for the purposes of this we just took the simplest path I mean again this is sort of the modus operandi is take the simplest possible thing and then see where you get with it so here we're
835
853
https://www.youtube.com/watch?v=P0yVuoATjzs&t=835s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
actually gonna generate whole frames so we actually want to confabulate and sort of imagine what the future frames gonna look like so you could imagine in that flash lag illusion example you're percept in that case is actually an imagined thing it's like it you're actually what you're perceiving at a given moment is is a fabricated future State so so we're gonna take it at the
853
872
https://www.youtube.com/watch?v=P0yVuoATjzs&t=853s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
pixel level but I agree because if you were interests in different things you might you might use different sort of targets for your prediction but what we did basically was to take the sort of classic idea of predictive coding and a state instantiated in the simplest possible way we could with with deep networks so basically what we do whose inputs come in here we have a running
872
891
https://www.youtube.com/watch?v=P0yVuoATjzs&t=872s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
prediction of what the input should look like and then we subtract them off to get an error map and then we send that forward to the next layer and then we have a recurrent layer at each at each stage of the hierarchy that's trying to build up this prediction and it can get feedback just the same as in the brain and it can also get local recurrence and then you basically just subtract off and
891
910
https://www.youtube.com/watch?v=P0yVuoATjzs&t=891s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
there's different ways you can do the subtraction doesn't much matter and then you get these errors you only send the errors forward through the network so one nice thing about this is at time one this is a CNN as nothing's come through so the extent that CNN's are a good model for the ventral visual pathway this is the CNN on the first time step and then the other thing that's nice
910
926
https://www.youtube.com/watch?v=P0yVuoATjzs&t=910s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
about it is it's also a classic generative model so if we put something in on the top it'll render down into an actual predicted image here so we can see what the network is is seeing or perceiving and I'm given moment and we called we started off by calling these particular works and then we don't maybe deep prediction that's not good name pretty deep no bill net no
926
946
https://www.youtube.com/watch?v=P0yVuoATjzs&t=926s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
can't believe it's not Alex net we ended up with pregnant because their rules their rules to how you do this and you have to you know you have to get them right so we're calling these pred nets for better or for worse and what we found is that these networks that had the recurrence at every layer we're able to you know our previous ones were only able to get sort of one degree
946
964
https://www.youtube.com/watch?v=P0yVuoATjzs&t=946s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
of freedom and a rotating object now we could actually get as many degrees of freedom basically as we wanted in these synthetic objects so here are examples and basically that just for all these I'm gonna show is this is the actual sequence of incoming images and then what's next but below it is the time-shifted prediction so you can compare the prediction to what actually
964
981
https://www.youtube.com/watch?v=P0yVuoATjzs&t=964s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
came and you can see on the first frame you get these weird potato thinks as it doesn't know which way it's going to go but then once it knows which way it's gonna go it locks on and then you start getting pretty good predictions and it can do this with faces it hasn't seen before and all that kind of stuff but again the the reason we're doing this is not for future frame prediction per se
981
998
https://www.youtube.com/watch?v=P0yVuoATjzs&t=981s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
but to see if we can actually learn good representations because the idea is in order to predict what's coming next you have to implicitly know lots of things about the structure of the object how light works how shadows work all that kind of stuff and what we find is basically if we try and build decoders again in the same sort of spirit as that face recognition
998
1,014
https://www.youtube.com/watch?v=P0yVuoATjzs&t=998s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
task where we just sort of peel off the representation and just do linear decoding we discover that we can decode lots of different parameters of the image like how fast it's moving what the starting angle is we can also look at things about principal components of the identity so you it sort of fits with this idea again that by virtue of learning good predictions you learn good
1,014
1,031
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1014s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
representations that are useful for lots of things which is kind of what you need to have in an unsupervised or a semi-supervised learning sort of setting we also did a face-recognition version of this again and what we found is that these presents at least at the time were performing on par better than the best semi-supervised learning algorithms that were available called latter networks so
1,031
1,049
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1031s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
you might wonder what happens if we put in things that aren't faces what does it do it actually does something pretty reasonable so this is a network that was trained on faces and we put in this image of a top and you know something came out pretty okay it doesn't always turn out okay here's a here's like a little toy car and you can see it's desperately trying to turn it
1,049
1,068
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1049s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
into some kind of face it's like this like when scream kind trying to happen there but you know and but the bottom line is if you train it on complicated things and enough variety you can get networks that can do pretty good predictions on just about anything you want to do and of course just about anything you want to do is cars right autonomous cars like at least in this
1,068
1,086
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1068s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
year we did this work that's the thing you want to be doing prediction on because that's where all the money is so we took my student bill without any prompting took some some car mounted camera datasets so this we train it on the kiddy data set which is in Germany and then what I'm showing you here is the test set which is the Caltech pedestrian data set and then this is
1,086
1,106
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1086s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
what the predictions look like so they're not perfect they're a little bit blurry if we put a Gann on it we can make it less blurry sure but we're really just focusing on sort of learning what kind of representations we can pull out of this and a couple interesting things here one is it seems to implicitly know a lot about perspective and flow it knows that things need to expand
1,106
1,125
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1106s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
outwards and needs you knows the things that are far away need to you know change less it also knows to infill road here if we look at other examples as well you can see things like you know it knows a little bit about occlusion so this car is going to include this golf cart it knows that that one should go in front it knows you know it knows all kinds of things about you know
1,125
1,147
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1125s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
implicitly about the flow of content in the image and again what we're trying to do here it's interesting to do it to look at feature frame prediction and we can also do farther out future frame predictions as well if we can go out as much as five frames it gets blurrier but it's it's still okay but again the goal here is to say well by virtue of learning how to predict as sort of a
1,147
1,170
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1147s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
surrogate loss can we learn how to decode other things that might be useful and in a car one of the things you might want to learn about is what's the steering angle of the car and what we find again in the same way we can take the representation peel that off put it into a linear decoder and we can discover that without any prior training on these steering angles this can
1,170
1,188
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1170s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
actually outperform a system that was purpose-built to decode steering angles so this is from comma this is a startup that's doing autonomous cars they had a reference CNN on their data set and then basically the president just by learning how to predict the future also implicitly learned what the steering angle of the car was that was kind of what you need to do to
1,188
1,207
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1188s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
do the task yeah not one not just the recurrency but also the fact that you're training it on continuously sample images representation is due to the prediction as opposed to merely the sequence of training images all right so so we have so we've done predictive versions and we've also done sort of a just a autoencoder style and the prediction outperforms the details
1,207
1,231
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1207s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
about are all in the paper how much additional how much of a better representation of you gain by adding the predictive component as opposed to just having from whatever image your sequence of images you have you mean specifically for this the steering angle version so the way the the comma one works is it it takes in frames and it puts them to a CNN and it so it has the temporal
1,231
1,258
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1231s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
component but it doesn't have any notion of prediction it wasn't it wasn't sort of trained to do prediction and what we see here is these were the comma reference CNN's so how well they performed and then this is with different numbers of input frames to the decoder and you can see it can can do quite well with with relatively little data so it seems to be a useful thing
1,258
1,277
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1258s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
but again what we're trying to do is trying to drive towards how might this be a principle that the brain could use to organize itself yeah when we're doing rollouts what do you mean oh into the future so their unique way school what we do is we re inject recurrent recursively the the predictions and then we fine-tune that so so basically if you take the prediction put it in again get the next
1,277
1,308
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1277s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
prediction put it in again put the next prediction and you can get its it starts to get blurrier and blurrier over time but that's basically what we're doing very briefly showed we can do five frames ahead the predictions just get murkier it makes a little bit of a difference it gets murkier because you don't know which way the car is going to go I mean realistic what we should be doing
1,308
1,336
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1308s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
what we are doing now is probabilistically we should actually be putting out a probability distribution of all the possible outcomes and that's that's hard but that's what we're doing and you know there's a lot of interesting sort of wrinkles but even just with this but the straight up let's just does this come up with a mean outcome because there's a problem that
1,336
1,353
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1336s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
if you don't know it's gonna go left or right then you you don't want to split the difference and get blurry what you'd really like to do is but have a system that produces samples that are sort of the correct distribution of actual outputs you could have and again that's something something we're working on something brain publi does - of course then we want to actually since this is a
1,353
1,372
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1353s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
neuroscience talk we want to go back and look at like well does this do something that explains something in neuroscience that we didn't understand before well here's something that everything does so it turns out you get good bores everything you know every neuron that work you train on anything no matter what you do it's like a rule that you'll get Gabor is out and and orientation
1,372
1,391
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1372s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
tuning but more interesting than that is this notion that I sort of flagged earlier which is if you put in a static input the brain doesn't give you a static output it gives you a dynamic output so there's usually a delay there's a burst of activity and then you have this sort of activity that falls off and sometimes you also get off responses so when the stimulus goes away
1,391
1,410
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1391s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
you get a fresh response and it's probably not a great surprise but pride nets do this as well and it's not hard to see why so this is the average of units in the error representation of the pregnant hard to figure out why it's doing this basically what happens is on that first hit it can't predict anything when the first image first flash is on so you get a big bunch of activity but then as it
1,410
1,430
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1410s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
locks on and learns to sort of explain away the data then it goes down and then when you get the image off that's another sort of surprise and then you get another burst of activity yeah if you sprinkled the pixels randomly like you know Toby Delbert's event cameras or something true continue so so it's true we're doing this in discrete time because we're talking about worker L
1,430
1,457
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1430s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
STM's and things like that convalesced hams so we can't accommodate continuous time but I don't think that's I don't think that's a huge huge difference well I mean these aren't like different frames it's like continuous frames so it's a relatively smooth progression through the space of images and when you have nothing and then you put on something that is the kind of
1,457
1,490
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1457s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
discontinuity we're talking about it is in discrete time yeah so what we did is we literally put the image up you know blank screen put the image up and then take the image off and so that's like the standard sort of primate experiment version of this and basically the present that does exactly the same thing has exactly the same dynamics that you would find in in primary visual cortex
1,490
1,511
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1490s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
so this is this is a bit of circumstantial evidence at least that part sort of matches up with expectations originally actually particular coding was it was designed to explain a phenomenon called N stopping which is if you have a bar that's oriented the way a v1 cell likes and then you make the bar longer and longer and longer and longer the response will go up to some point and then once you
1,511
1,534
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1511s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
make it longer actually the response gets suppressed so it's like too much of a good thing you know the longer the bar is that that's in the orientation that likes it actually reduces the response and you know true to form the the present version of predictive coding also has this sort of n stopping response which which is a nice nice thing to have but it goes further than
1,534
1,551
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1534s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
this there's also this notion of surround depression so if you have a stimulus which is a dot and you larger and larger and larger and larger the e1 cells will like it better and better and better and respond more and more and more up to a point but then beyond that they'll start being suppressed as if there's a suppressive surround around the response and what we
1,551
1,566
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1551s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
find is that the the pet net also has this quality so if you look in again the bottom layer the response goes up to a certain size of a dot stimulus and then it's suppressed but interestingly Rick borns lab at Harvard back in 2013 found that if you cool downstream visual areas so if you cool v2 and then record in v1 so in activate feedback connections you actually would find that these surround
1,566
1,591
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1566s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
suppression effects were themselves suppressed they go away if you take away the top-down feedback and this is a way easier experiment to do in in a network because you can just turn off the feedback connections and see what happens and lo and behold you get almost exactly the same pattern of feedback so in the pred net sit extent there surround suppression it's happening as a
1,591
1,610
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1591s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
top-down feedback phenomenon just the same way it seems to be happening in the primate and then you know as we sort of march through these things we also have this interesting sort of sequence learning effects that you see in the end of the ventral visual pathway and lo and behold if you show it pregnant these things so the pred Nets just trained on natural images like our videos and then
1,610
1,633
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1610s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
we're gonna show it these stimuli that are basically the same as what you would find in these experiments in an untrained pred net you get these sort of funny effects where it's you know you flash up an image and it takes a little while to sort of predict it away but then as you train these these sequences so that are expected you get sharper and sharper transitions from between these
1,633
1,654
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1633s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
images and then what happens is basically the same exact result so if B is predictable from a in serial presentations again this is trained on cars and then we just subsequently show it a few of the you know a comparable number hundreds of examples then B will be suppressed if a explains away B but if we show that very same B and you know with something else that's not a that
1,654
1,676
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1654s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
doesn't predict it in front of it we get the exact same result that the Carlos and company got so it's able to track with almost exactly the same number of trials some of these interesting sequence learning facts with no no extra machinery needed to be added and then you might also ask okay that that flash lag illusion so you might remember you know this this clocks
1,676
1,697
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1676s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
going around this bar is flashing it's actually lined up but it looks like it's lagging behind so you might ask oh is that this is the same thing happen with with the pred net and the cool thing about generative neural networks is you can see what they're experiencing because we can just visualize the error layer and so this is what it's seeing now there's a caveat here does anyone
1,697
1,717
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1697s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
know the caveat is yeah yeah so you're getting a double dose so so the network is seeing it and then you're seeing it on top of what the network is seen so if we really want to do this we need to we need to look at freeze frames and we can see exactly what what the network's seeing and there it is there's that weird lagging and there's the tilt too right that was weird
1,717
1,736
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1717s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
there wasn't a good there wasn't a satisfy satisfy explanation for that tilt before and it just sort of falls out of the pregnant yeah just predictable on a longer time scale yeah so it's the same thing with with with the human version of it right like it's predictable but there seems to be some window beyond which it's not it's effectively not predictable that's right
1,736
1,762
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1736s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
there's also weird things like you see these like weird ripoli things happening and I don't know what those are but they're really cool and I like looking at them and we'll figure out what they're for eventually and then I love the scientific community and I love archive so we posted this and a group in Japan picked it up and started using it so they basically wanted to see these
1,762
1,783
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1762s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
illusory motion stimuli they wanted to see what would happen if they showed them to pred nuts so so the way these work hopefully you're experiencing the illusion right now this is a static image but it looks like these are are sort of these ones are sort of are sort of rotating and these ones are not so this is a classic you know sort of illusory motion sort of stimulus and it
1,783
1,807
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1783s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
turns out if you if you train a pret net on these sort of rotating so it's seen real objects rotating in the world or synthetic images of like propellers and things moving in the world and then you compute the flow vectors on the actual predictions if you just let a pregnant look at these static images the pregnant actually produces flow vectors that are consistent with the illusion so when
1,807
1,826
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1807s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
humans see the illusion the pregnant has optic flow that looks like what we see and when humans don't see the illusion there's no optic flow so so this is this is interesting especially because the usual explanation for these things has to do with sort of epiphenomena about the relative Layton sees visual responses and now that might still be true but this at least gives you another
1,826
1,848
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1826s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
possible explanation which is that you have a system that's trained with the predictive loss it's gonna naturally have some of these biases that come from the sort of statistics of the world so with all things neural network you know like we just hide all the like the parameter search that went into it there is indeed so the failure mode in general when it can't when it doesn't learn so
1,848
1,881
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1848s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
if we have like the hyper parameters wrong or whatever is that it predicts the last frame that's usually our benchmark against which to see if it's working so we take the reconstruction lost if you just said it's the same as the last frame and then compare how much we improve the error and then you're right it if it's too herky-jerky then it has trouble learning so so I don't have
1,881
1,899
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1881s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
a good quantitative answer for you but there there is indeed a sweet spot and yeah I mean it depends on how fast you run the framerate we were running these at 10 frame frame rates so that's kind of yeah about about 100 milliseconds it depends on how fast things are moving as well so anyway so I promised that we would get back to a neuroscience experiment so I'm
1,899
1,922
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1899s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
going to tell you about a story about an actual neuroscience experiment that the neuroscience said the wet lab part of my lab did in response to the computational work that was going on in my lab so you might remember that we could read out the steering angle from a self-driving car using the internal representations in the pregnant we also had the idea well you know wouldn't it make sense to
1,922
1,942
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1922s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
actually take the efference copy of the steering wheel all the odometry of the car we could just feed that into the representation surely the network would do a better job of predicting so if I'm trying to predict how the world is going to change if I know the wheels turn this way I can make a better prediction about how the world's going to change I might also be able to do cool things like if I
1,942
1,959
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1942s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
could sort of fictive Li imagine like conditionally generate if I turn the wheel like this what would the world look like so this is something we started doing just because it made sense from a machine learning standpoint and no big surprise when you don't have a friend's copy which is basically the sort of vanilla pret net you get certain convergence over time you know so you
1,959
1,979
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1959s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
get get down to some mean error and with with training epoch but if you include this extra information unsurprisingly the network converges faster and comes up with a better result so that's that's all fine and good but if this were true and this is what was happening in the brain that would imply that in visual cortex we should have signals from motor cortex right they should be there and if
1,979
2,001
https://www.youtube.com/watch?v=P0yVuoATjzs&t=1979s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
they're there we should be able to decode them even in the dark potentially so so my student Greg thought this was a great idea he was doing 24/7 recordings in visual cortex in a rat at the time so he's had boatloads of boatloads of data so we just had that lab meeting well hey well I just put them in a dark box like so there's no light completely light-tight and see if you can actually
2,001
2,022
https://www.youtube.com/watch?v=P0yVuoATjzs&t=2001s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg
P0yVuoATjzs
and you have an accelerometer on the animal's head anyway because you know the minute you're putting electrodes and you might as well put an accelerometer in there too and then so we can record from from many tête roads so we have you know 60 16 tetrode 64 electrodes and we can record local field potentials and then we also have these accelerometer signals so there's a
2,022
2,042
https://www.youtube.com/watch?v=P0yVuoATjzs&t=2022s
Predictive Coding Models of Perception
https://i.ytimg.com/vi/P…axresdefault.jpg