ICLR2015-terry-sejnowski-part1

What this episode is about

The discussion emphasizes the integration of neural networks and biological learning algorithms, exploring the parallels between artificial intelligence and the central nervous system across varying spatial scales. It highlights the necessity of situating deep learning networks within a real-world operating system, stressing that biological networks operate in an environment shaped by evolution and behavior rather than controlled conditions.

Key takeaways

• In a drought, washing outside might just become the hottest new trend in San Diego.
• Deep learning networks are like pampered pets—they need real-world behavior to thrive, not just coddling.
• Our brains, honed by millions of years of evolution, remind us that complexity often begets simplicity.
• More isn’t just better; it's evolutionary—nature's secret moves in layers, just like your favorite deep learning model.
• The hippocampus is like your brain's feedback loop; it's the ultimate connector in your sensory hierarchy.

Who speaks in this episode?

• Terry Sejnowski (Host) · 30 min speaking

What is discussed, and when

1. 0:00 – 5:00: Theodore Djony
2. 5:00 – 10:00: Hippocampus, Paul Glimer
3. 10:00 – 15:00: Toby Delbrook, Natural Images
4. 15:00 – 20:00: Natural Images, Percolation Theory
5. 20:00 – 25:00: Bob DeSimone, Ken Wilson, Natural Images
6. 25:00 – 30:00: Temporal Difference Learning
7. 30:00 – 33:28: Hippocampus, Jeff Hinton

Notable quotes from "ICLR2015-terry-sejnowski-part1"

• It HC okay that is not the homunculus that is the hippocampus now not it's not just the visual system but it turns out there are similar hierarchies in the auditory system Thea sensory system every major.

  about Hippocampus · at 4:58

• Information coming in from the retina or an image or a movie right well here is an experiment that was done by Paul glimer this is over 10 years ago which suggested that it's…

  about Paul Glimer · at 6:05

• Simple feature on and off so I'm going to show you how far you can get with it so this is a camera that was developed by Toby delbrook it's a neuromorphic camera it's built he…

  about Toby Delbrook · at 11:31

• Designed to do is to be able to pick out the structure from these complex scenes one of the properties that we know about natural images is they scale invariant so what I'm…

  about Natural Images · at 15:11

• And he applied a u theory in physics which is percolation Theory that you may not be familiar with the idea is that you have a grid and that the probability of any one of these…

  about Percolation Theory · at 16:20

• In condensed matter Theory phase TR Transitions and very powerful theory behind it Ken Wilson has a whole Theory got Noble Prize for called the renormalization group and here…

  about Ken Wilson · at 19:55

• That I showed you earlier and what he noticed was that if you count the spikes that on the second trial it's almost always less but there were a few cells where it increased…

  about Bob DeSimone · at 24:15

• Release and still function have a functioning brain well I know the answer to that in fact Jeff H Hinton told me how the brain works so I know so this is from a lecture that Jeff gave which is he was given a prize the kilum prize in.

  about Jeff Hinton · at 33:01

Who and what is mentioned in this episode?

People

• Toby Delbrook — Person · Referenced technologist (device creator) · 1 mention · first at 11:37
  Credited with developing a neuromorphic event-based camera used as an example of efficient spike-based sensing.
• Bob DeSimone — Person · Referenced neuroscientist (hypothesis originator) · 1 mention · first at 24:08
  Introduced as the source of a hypothesis explaining repetition suppression/perceptual priming via sharpening of neural representations.
• Paul Glimer — Person · Referenced researcher (study author) · 1 mention · first at 6:10
  Cited for a monkey vision/reward experiment supporting the claim that visual cortex encodes value, not just stimulus features.
• Ken Wilson — Person · Referenced theorist (renormalization group) · 1 mention · first at 20:01
  Cited to connect natural-image statistics and criticality to renormalization group theory and scaling ideas relevant to deep learning.
• Jeff Hinton — Person · Referenced expert (anecdote/authority) · 1 mention · first at 33:05
  Invoked as an authority figure in machine learning; speaker jokes that Hinton ‘told me how the brain works’ while discussing noisy synapses vs precise weights.
• Theodore Djony — Person · Referenced expert (attributed quote) · 1 mention · first at 1:58
  Credited with an evolution quote used to frame the talk's biology-first perspective on learning.
  System okay well nothing in biology makes any sense except in the evolution that was a quote from Theodore djony evolutionist and what I've Illustrated here is the U spiral of…

Concepts & topics

• Hippocampus — Concept · Core topic · 3 mentions · first at 5:07
  Highlighted as essential for long-term memory and learning unique objects and events, with detailed discussion on its structure and function.
• Basal Ganglia — Concept · Key argument · 7 mentions
  Discussed as a critical part of the brain involved in decision-making and learning, particularly in the context of reinforcement learning.
• Natural Images — Concept · Strategy · 3 mentions · first at 14:45
  Used as an example to discuss the properties of images that the visual system is tuned to process.
• Percolation Theory — Concept · Framework · 1 mention · first at 16:20
  Applied to analyze natural images, suggesting a phase transition in image bit planes that mirrors physical systems.
• Temporal Difference Learning — Concept · Framework · 1 mention · first at 27:36
  Discussed as a learning algorithm important for classical conditioning and dependent on dopamine.
  Conditioning by Reed monteu and Peter Dean when they were post in my lab depends on dopamine dop here's here's the world which is being represented by the cortex and the idea…

Full transcript with timestamps

0:00 first of all want to welcome everybody here to sunny Southern California I have to alert you that we are in an extreme drought condition and the mayor of San Diego recommends that if you need to take a shower you should take off your clothes and walk outside so this lecture is uh actually this meeting reminds me a lot about a lot of the uh early snowboard meetings in the uh 8586 era before nips I mean it has that kind of a Revival Camp feeling uh which is very exciting because you know you're sort of on the cusp of something not quite sure what it's going to turn out to be U so this lecture is going to have two themes one of them has to do with putting networks and and and if this levels diagram from molecules up to central nervous system well which spans you know nine orders of magnitude of spatial scale you know the uh the the networks there's really only one band in the middle of that and

1:04 there's a tremendous amount of Machinery down here at the level of synapses and that's one of the focuses of the talk today is that what do we know from nature about um learning algorithms that that might be relevant to what you're trying to do and secondly embedding these networks into an operating system I mean you need to have something around it right now these uh deep learning networks are coddled you are providing this the sensory input you are providing the training data you are providing everything it's like a little coddled little network but you know our brains work in the real world I mean there isn't any uh way to uh access it except through uh behavior and so that's really the question is how do you uh wrap the networks around uh the operating system okay well nothing in in biology makes any sense except in the evolution that was a quote from from Theodore djony evolutionist and what

2:03 I've Illustrated here is the U spiral of life and as you can see start starting you know three I can't read it three billion years ago um we progress through the bacteria through uh UK carots of fish and eventually uh mammals and so the mammals have only been around for about 100 million years uh primates for you know maybe 20 million years man has only been here around for a few million years so in a sense we're the end point of a lot of previous Evolution and uh we can learn a lot by comparing comparative U Anatomy so if you look at these earlier creatures they didn't have a neocortex the neocortex with all of the you know the hierarchy that's all Maman uh and what distinguishes it from the three cortex that you find in um in for example U reptiles and birds uh is that uh those networks were highly associative that is to say they there was no topography uh the Maman six layer cortex is mapped the

3:17 visual input from the retina is mapped to the cortex and remapped over and over again and it's that Innovation that has made it possible to go from a very simple cortex that is as you can see here in stipple gray for these early species uh Hedgehog and then a prmium galago and then finally human you can see there's a tremendous expansion of the stable cortex more is better and it's very rare that something scales this well that's really the secret of evolution is it discovered something that is you can reproduce relatively easily and throw into sensory systems and other Association systems and so that's what really what I want to talk about so this is the famous venessa diagram of the visual system Circa 1990 starting with the retina here at the bottom and as you can see each one of these boxes represents one of these maps in the visual cortex uh at the primary visual cortex V1 but there more than two

4:21 dozen other areas and you can see that this subway map here connects them and you can see that uh this resembles to a very large degree what we now call deep learning right we have uh many many levels and we have representations that get more and more complex the the receptive fields of the neurons that say the region over the visual field where they respond gets bigger and bigger until finally you get up to the infratemporal cortex where you recognize objects and you learn objects and this is really the level at which we're we're interested right we're interested in objects now you can see that living above the visual cortex is a little box I don't know if you can see it HC okay that is not the homunculus that is the hippocampus now not it's not just the visual system but it turns out there are similar hierarchies in the auditory system Thea sensory system every major sensory system funnel into the uh I was

5:17 going to say homunculus uh up here and and it's it's that part of the brain which is responsible for learning unique objects and events so that you can code ify your experience and and use that in order to guide your actions more than just the semantic knowledge that you have down here so one thing to think about is you know well what are the outputs uh the thing about the hippocampus is its only output is actually feedback into this hierarchy in fact every single forward connection has a feedback connection so this is going up and down and that's also a very important principle okay well visual system it you know the tradition is that you have uh if you're trying to solve a vision problem what you're trying to do is make a model of the world based on the information coming in from the retina or an image or a movie right well here is an experiment that was done by Paul glimer this is over 10

6:16 years ago which suggested that it's there's a lot more going on in the visual system than than we thought so this is a simple experiment the monkey is fixating the spot and it is instructed on different blocks either to move its eye to the right or to the left now what the only difference between the blocks is that the amount of juice that the monkey is going to get varies with the position so in Block one you get twice as much juice to the right and block two is twice as much to the left now we know that this part of the brain that there are neurons that are selective for both the visual stimulus and for the action that you're going to take in this case moving your eye to the right so a visual response as shown here this is a spike firing rate as a function of time and you can see that in response to uh the light coming on the the monkey moves its eye and uh you get a visual response and then a little bit

7:13 later uh well first the visual response and a little bit later the monkey actually targets the uh that point now here's here's the amazing thing is that you get three times more spikes when the monkey is making a cicad to the to the Target that has more reward than to the Target that has less reward in other words these neurons are encoding a lot more than just the visual Target they are encoding the value of the Target and that is a completely different ball game right because you're saying that you know this is not a vision box this is a vision evaluation box that we dealing with here so that really is is change everybody's view about Vision here's another uh output from the cortex this is a part of the brain called the super calculus and it's important for targeting eye eye movements and this is the parial cortex projects down to it and but the parial cortex is only one of two dozen different areas and this is

8:13 from the cap Rin each one of these little slices is from a different cortical area showing the termination of the inputs in different layers so the spiculus has many layers uh superficial which are sensory and deep layers which are motor but you can see that there's a tremendous convergence all the areas are converging even the earlier ones in other words moleculus is picking up information from wherever it can get it and and it's purpose is to help Orient the animal towards something interesting which could be either a bright flash in which case V1 would be a good Target input to to focus on or it could be something coming from the prefrontal cortex down here which is something that you're looking for you're not being driven by the sensory input you're being driven by some internal state so this is another lesson is that you shouldn't just look at the output layer for information there's information at every

9:05 Point all the way up the ladder it's very important if you're trying to make a quick decision now in addition uh so there are the subcortical motor targets but what about um uh and we're going to come back to this reinforcement learning so it turns out the entire cortical mantle projects down to uh the basil ganglia and specifically to the striatum uh and and what's shown here is three different parts of the cortex there is um on the very uh left is the sensory motor this is the part that is helping you make decisions about how to move your body around and and you can see that it projects to the red regions here of the basil ganglia and interestingly projects down and it comes back through the thalamus and projects back to the cortex there's a loop and and that Loop has a delay of about 100 milliseconds and it's thought that the basil ganglia actually where converging information is compared and decisions

9:59 are made about actions but uh you have similar parallel Pathways from another part of the cortex the this is the dorsal lateral prefrontal cortex that has yet a different subset of the basil gangle and then yet another part which I'll I'll come back to which has to do with making important decisions for the future planning so these are not all equal in terms of their function well I want to start at the beginning in the retina I make illustrate an important point about representation before moving on to these uh other issues so this is a mouse retina and what you see here on the right is a schematic and this is a real retina that's been stained photo receptors at the top some processing uh through these two layers and then finally the gangion cells which project into the brain which are shown here in kind of a what would you call that purplish mve or something um so interestingly the light comes from

10:54 the bottom and it goes through all these neurons before it gets to the photo receptors go figure the gangling cells uh there are about 24 different types of gangling cells that have been discovered in different um Maman retina and the two prominent uh types are called the uh on Center and offc center and the on center uh provide input when the light comes on in the center region little circle and the off provide a burst of spikes when the light goes off so that this is telling you something about change in light intensity now what can you do with that simple feature on and off so I'm going to show you how far you can get with it so this is a camera that was developed by Toby delbrook it's a neuromorphic camera it's built uh he built the Chip And and as you can see uh here's the form factor the thing weighs you know 10 gram it's uh consumes microwatts it's incredibly efficient and here's all it's

11:54 doing every pixel in the camera has two different channels there's a channel that produces a spike when there's a small increment and there's another channel that produces a spike when there's a small decrement so these are the on and the off and so it encodes a continuous waveform in a sequence of pulses that it gets transmitted as you'll see uh into the into the like the brain uh into the uh computer let's see if this is going to work oh movie is not going to work but uh what what you can let me uh quickly scroll down here and bring up this is a long

12:57 talk okay uh okay so I'll just describe what the um the output of the camera it's too bad because it's a really it's a wonderful uh illustration it speaks for itself so U what it what it what it shows is that when there's nothing moving there's no output it's blank and and and it's only when something moves and when something moves you can see on the Leading Edge little white spikes and then on the trailing Edge little black spikes now the beauty is that with this camera you can follow a rotating spot that's that's going at 200 HZ 5 milliseconds per Revolution if you took a picture of that with your frame base camera it would be a circle right because the frames are like 30 milliseconds but with his camera you can actually follow with a microsc Precision the leading and trailing Edge and from that you can compute the velocity so the point is that if you can preserve the information at the earliest stage terms

13:59 of extracting the features that it's going to be a lot easier Upstream to be able to compute things like velocity stereo and and this is uh one of the virtues of using spikes is very efficient low power and this is actually equally important for what you doing because everybody here okay is is simulating things on digital computers that are very power hungry well ultimately you know you do want to be able to deliver this uh at a low power and so you're going to need a technology to do that so I think this is something to be thinking about into the future is that you really have to think about how to translate the current algorithms into a format or a a hardware that is going to be uh deliverable that'll be scalable and deliverable to robots for example okay I want to change gears a little bit here and talk about natural images so natural images is of course where we evolved uh in in the savannah and U and our visual system

14:55 is really tuned to the features in that environment millions and millions of years ago this is a database from a single image from a database showing you know a field and it has a lot of texture in it but you know where are the objects but this is what we our visual system was actually designed to do is to be able to pick out the structure from these complex scenes um one of the properties that we know about natural images is they scale invariant so what I'm doing here is down sampling on successive frames here and you can see that even at the crudest level here where you can see the pixelated image there's structure there's interesting uh gradients and uh which reflects the actual uh information in the scene and what this illustrates is that no matter what scale you're looking at uh there's going to be something that uh you can extract and if you do a fer analysis a power Spectrum this in this case the spatial

15:51 frequencies uh you get of one over F uh squar Power and and that's a that is a sign for power law scaling so this is something kind of a mystery is that uh you know what is this what's going on in nature that uh and is this mirrored somehow in in the brain so uh this is a project uh now I'm going to tell you very beautifully about from u a postto myb s Sami who's here uh background in condensed matter physics and he applied a u theory in physics which is percolation Theory uh that you may not be familiar with the idea is that you have a grid and that the probability of any one of these uh little pixels here being on is is given that's the that's the probability p and they're being turned on and off randomly you just pick them at random and turn them on and off with that probability and and as you increase the probability you go from uh this uh a very low probability event where you get small little Islands to uh for large

16:53 probabilities the whole thing is going to be filled but there is a point right here in the Middle where Something Magic happens and that's when these islands get large enough so that they can actually completely span top to bottom and that's the percolation threshold and that turns out to be a phase transition in physics between two different states the state of being blocked and the state of always having a place to get through and this has lots of applications as you can imagine in in in in uh if you're uh into trying to find um uh whether soil is going to pass water turns out you can also apply to images so here's the the uh it's it's a novel U representation of the pixel intensity the gray level uh in terms of bit planes right so here's the highest high level pixel here and here's the lowle pixel here so there's uh in this case here there's 14 pixels the vanh hatan database has has I think 15 pixels of

17:50 resolution and so the idea is that uh what we can do is look at every plane uh here's the the image and here is uh the uh bit plane one and that's the top level and you can see most almost all of them are off that's because there isn't there very few places that have very high intensity now jumping to the bottom here here's bit plane uh 10 and as you can see it's complete noise because of the fact that you're dealing with fluctuations in the sampling and so forth and again somewhere in between around bit plane five or so as you can see here you can see that there's a tremendous amount of information about the original scene it's just in one bit plane in plus or minus maybe one or two around it well we can analyze uh the statistics uh and and here's what uh say found um what he found was that there's a percolation phase transition in the images at bit plane six around six you can see that all this is all a connected

18:49 set of pixels in red connecting the bottom to the top and if you're below that you get a lot if if you're uh if you're yeah if if you can see the uh in this case here uh this is these are bit planes that um have lower probability uh below it uh and this is a bit plane above it you can see that there's a lot more uh connections here and here's and this is now the uh the the probability of having a direct connection from top to bottom as a function of the bit plane and you can see that uh it it drops precipitously at at this point here and that turns out to be uh a to have a critical exponent here uh which is the power law Lambda C minus Lambda C is the critical bit plane um with a universal exponent of 6 so with this I I you know this I can't give a whole lecture on condensed matter but this is an amazing Confluence where it looks as if analyzing images of Nature has a lot of the Hallmarks of what what what's called

19:55 in condensed matter Theory uh phase TR Transitions and very powerful theory behind it Ken Wilson has a whole Theory got Noble Prize for called the renormalization group and here here very briefly is what it looks like if if if you take and down sample an image like this uh and you end up if you're right at the critical point you end up with a uh a small number of pixels which has the same statistics as as the large scale one that is to say there are correlations at all length scales that's what we sh Show that's what we showed in the uh natural images but if you're uh if you're above that uh you can see that uh you you will either wash out uh for all zeros or all ones if you're above or below and what this suggests is that uh knowing something about where you are here in terms of where what what input you're going to give to the to the to the Imaging system is really critically important 90% or more of the information

20:52 is contained in that one plane image plane plus or minus one bit it was a paper that came out last year by meta and Schwab in archive which was very interesting making pointing out there's a formal analogy here between what's going on in renormalization Group Theory which is um decimation where you are averaging over a plaque of pixels and then assigning it plus or minus one depending on the majority and what's going on in deep learning when you have an architecture where the units on the top are seeing more and more of the inputs as you go up so this this is this is really interesting the suggestion here is that maybe part of what's going on this is restricted Boler machines may be that you're at a critical point and that's something that ought to be looked at more carefully okay now in the remaining part of the lecture what I want to do is to convince you that there is much more to do uh in trying to

21:48 understand learning than back prop or boltzman learning okay the brain Nature has actually evolved dozens and dozens forms of synaptic plasticity that are very important because they work together synergistically and I'm going to show you a few of them so priming so priming is something psychologists have studied for a long time it's a form of one trial learning it's unconscious you're not aware of it but it has very lasting effects and so here's an example suppose I gave you this image and ask you identify what is that can anybody guess what this is I've erased half the lines what okay a couple of people guessed the right answer so it turns out that that uh if I show you this Frame first and then show you this then everybody will see it and that was the whole series of experiments have been done in fact with this particular uh experiment uh that was done on uh you know psych 101 uh and

22:52 you know you bring them back a week later and you show that uh they can fill in the blanks and the the the the the investigator actually 17 years later tracked down those students sent them these images and found out they could still recognize them so it's it's it's incredibly powerful and it's one trial and it's always with you isn't that amazing it also works for uh words for example can anybody guess uh there i' I've erased a couple of the letters here where there are blanks and can anybody guess what this word is you showed us before okay I've already showed you this okay in that case it's cheating Okay now what's amazing about this is that in fmri when you look at this this the uh activation of the subject and this is looking at the brain horizontal View and this is the back and this is the front uh and if you look at the first presentation you get a tremendous amount of activity here in

23:47 the visual system second presentation you get much much less this is called perceptual priming and uh it's called repetition suppression so this is one trial so it means that something has been reorganized into your cortex after one trial now here is here is a a a theory a hypothesis that Bob DeSimone came up with in 19696 he was recording from the very top level in the inferior temporal cortex that I showed you earlier and what he noticed was that if if you count the spikes that on the second trial it's almost always less but there were a few cells where it increased majority decreased and here's this idea that the very first trial you get a few cells that are firing at high rate but most of them are firing at some intermediate medium rate and here's uh what he thinks happens is that there's learning that takes place that changes the strengths of the connections between the neurons that are firing at a high rate so that

24:44 they're increased and decreases the strength of the connections with the medium weight so that after the first trial now you you get a much more focused uh and sharpened representation in which you've released a lot of the units that otherwise would have participated in the in the representation so this is a form of sharpening and it's something that is is unsupervised it's happening all the time uh one of the posts of my lab mold carof has simulated this a two-level model early visual processing late visual processing occurs uh in the second layer and this is a winner take all and uh what he did is implemented that learning rule I just told you about and here's what happens in trial one you can see that a few of them are firing at a high rate most of them fire at a medium rate or low after that learning it takes place you can see that two things happen first uh these are suppressed so that their uh activity

25:39 level is much lower but you can see that for the ones that are firing at a high level they reach thresh they reach this high level much more quickly and it's also observed psychophysically that on the second trial you can respond much more quickly to the stimulus makes it more efficient you have a sparse representation here's another form of repres of of plasticity it's called homeostatic plasticity again this is unsupervised and the idea is that if a a neuron uh has a synaptic strength and for some reason is not firing has reduced activity all of the strengths are proportionately increased they're scaled up and if it uh there decreased activity if if there's increased activity the cell's firing too often then you scale down and the idea here is to maintain it within some operating range it's a very simple algorithm we know the mechanisms the molecular mechanisms it's all it's there in almost

26:26 all neurons well here's an interesting experiment that was done by Ying trsh uh and it's called the bars problem and the idea is that the inputs to the network are going to be either horizontal or vertical bars with a certain probability here are some inputs and uh this is a uh nonlinear independent component analysis problem that cannot be solved with normal I algorithms and you're trying to pick out the uh the the features which are the the bars and as you what he did here was to combine simple heavan plasticity between two layers and also a homeostatic mechanism in the upper layer and as you can see after many trials uh you begin to to form these uh very U uh individual uh features and interestingly the um the weights of become very sparse these are almost all small weights and there's only a few large ones so this suggests that you combine two different forms of plasticity on supervised in different in different you

27:26 know different types of neurons and you can solve these complex problems that otherwise a single algorithm might not be able to solve and finally and this is something that I I'm I'm pretty sure you're all familiar with I just want to point out that temporal difference learning that was introduced by Barton and stto in the 80s and was suggested to be important for um classical conditioning uh by Reed monteu and Peter Dean when they were post in my lab uh depends on on dopamine uh dop here's here's the uh the world which is being represented by the cortex and the idea is that uh the the basil gangling here is going to compute the value what's the value of that state and it's now going to use that to update the weights uh depending on the the the uh the feedback of the reward and that will also be used to make decisions so this is the traditional actor uh critic architecture uh well here's where the

28:20 basil gang are located in case you were wondering I'm not going to go over the details except to say that this is again it's a it's a dense projection converging from the cortex onto this little structure and then uh the dop me is really remarkable there are these dope nuclei down here in the midbrain that project up into the entire cortical mantle including the basil gangle very heavily inovated incredibly important for Addictive Behavior every addictive drug goes through uh increasing dopamine levels of activity now we come back to the hierarchy I was telling you about earlier so we have these uh three levels sensory motor we have the associative Network the this is the dorsal lateral prefrontal cortex and then we have the uh the limic cortex which is the uh the frontal pole and uh it's called the vental strum so it what's interesting is each one of these does something different this one Associates a sensory

29:13 input with an affector like an arm and it's it's it's very it's like a habit and it's very uh difficult it's not flexible so you can't generalize from one arm to the next you can't generalize from an arm to a foot if you wanted to have some kind of a um Fast Response the dorsal lateral prefontal cortex which contains a lot of working memory cells cells that can maintain activity for a long time is capable of associating a sensory input with an outcome so and that outcome can be flexible in terms of you know what arm you're going to use or what foot you're going to use and finally the very highest level here which is the orbital frontal cortex the dorsal this is the uh premotor cortex these are areas of the brain that are important for planning and they uh if you damage them you damage people's ability to do uh discount temporal discounting that means that they can't think about the future

30:07 they they they take things uh they're they're greedy uh and so so we we think we this is an area that psychologists have focused on for what's called prospective optimization so you want to optimize things so one way you know in terms of your memory you can look back in the past and you can use that information to plan the future and and prospectively pick the actions that are going to be valuable in the future which could be you know in the distant future so uh so and this is if you're all interested in following this up I have a paper in the proceedings the i e last year where uh together with Simon with several others we worked out the kind of the consequences of this kind of architecture for uh how you would design a system an operating system to be able to handle uncertain conditions and plan for the future so I've shown you the couple examples priming homeostatic plasticity temporal

31:04 differences these are all different learning algorithms and they're all occurring simultaneously at the same time uh and and it's very I think it's very clear that there's also going to be plasticity with dopamine proje of the cortex it's going to control representation learning in the cortex having to do with the value and I showed you an example of firing rate which depended on value one last thing I want to show you this is from a very recent study that I did uh a group of my lab Tom Bol and others was to look reconstruct the piece of the hippocampus the part of the brain that's important for long-term memory and it it's well established that these are spines the synapses around the spines and the red is the uh the active Zone where the transmitter is released and they come in a tremendous variety of sizes and shapes it turns out that the size the volume of that spine is proportional to the

31:51 synaptic strength and this is really amazing what we determine and I won't go into the details is that you store five bits of in the synaptic weight and we did it by looking at pairs of spines that were interated by the same axon that had the same input and had the same post synaptic voltage and they turn out to have the same size spines is that per synapse or per connection between neurons I mean per pair of neurons yeah per per synapse every synapse including the small ones have this five bits of accuracy now here's the surprise here's the amazing thing the amazing thing is that the probability of release only one vesicle is released the probability of release for the majority of the synapses is about a tenth there's only a few very big ones up here that have high probability and that means that literally eight or nine times out of 10 that you stimulate the nerve you only get you you get no

32:49 release tremendous amount of noise how would you have a synapse that has so uh High Precision in terms of the strength has have such a low probability of of release and still function have a functioning brain well I know the answer to that in fact Jeff H Hinton told me how the brain works so I know so this is uh from a lecture that uh that Jeff gave which is uh he was given a prize the kilum uh prize in Canada one of the highest prizes for a scientist and he gave a lecture on as you can see brain

Questions about this episode

Who speaks in "ICLR2015-terry-sejnowski-part1"?

Terry Sejnowski (Host).

What is "ICLR2015-terry-sejnowski-part1" about?

The discussion emphasizes the integration of neural networks and biological learning algorithms, exploring the parallels between artificial intelligence and the central nervous system across varying spatial scales. It highlights the necessity of situating deep learning networks within a real-world operating system, stressing that biological networks operate in an environment shaped by evolution…

What is mentioned in this ICLR episode?

11 people, works and ideas are mentioned, including Hippocampus, Basal Ganglia, Toby Delbrook, Bob DeSimone, Natural Images, Paul Glimer. Each one on this page links to its own page with every episode that mentions it.

How long is "ICLR2015-terry-sejnowski-part1"?

The episode runs 33 min; the transcript is about 5,943 words in 35 timestamped sections.

More from ICLR

• All analyzed episodes
• People mentioned on ICLR
• Topics mentioned on ICLR