 και μειάζει ότι είμαστε σχέσεις. Γεια σας, καλώς επειδή και ευχαριστώ για μια άλλη σύντομη της Σασάσεξ-Βυζιονσέμμινας, όπως οδηγιόμαστε στην ΑΕΕΚ, και αυτή η ώρα συμβεί με το ΛΕΛΕΙΕΕΚ, από την αυσιασμότητα και τη προστασία για την αυσιασμότητα. Είμαι ο Γιωρτς Καφετζής, ο μάστος τελευταία από το Τόμα Σοηλίσσο και τώρα, παράδικα με τον Μπαδεν και σκοτή σας, ευχαριστώ να πω πάνω, ήταν νερό, για να βρήθω ένα εξηγένειο γύρω από την Παρικογένση ασ Sebinar world, και να συμβαίνω ανθρώπη με τη φράση του αντισπίστρατοι, ο Μαρκούς Μαίστρ. Μαρκούς συζητήσε τη φιση στο Τεχνώνη Ουνιυένου, Βεβαίνει σε ΗΩΕ και σε Καλτεκ στην ΕΕΖΙΤ ΕΖΙΟ ΒΉΛΕ腳επλαμβανόκητου σε την συνέσθηση της Αμα vérωσης του ίΡΑΜΑΑΚ. Σε από τα δακτάκο να τ αναπτήρει που η ΕΚΟΡΙΤΙΟ ΥΠΑΙΣΑΔΑΙΑ είναι να σκέψει για την ευρώπη, στο 2012, που το συμβουλείται για την ΕΚΟΡΙΟΙΑ, στο οποίο έχει been located ever since και πιglοντικά να αρχει το κειφ, και Μπενζαμμιν Γιαγγίνη, Προφέσσο της Βιολογικής Ζιένσης. Στην μάχη του μόνος, ήταν σαν την πόστα του Γιας στον Στανφορ, που διεθνήσε τον Μπενζαμμιν Γιένης, που εξεγραφήθηκε στην Βιολογικής Ζιένσης και την μονοσυκλέτα της Ρετίνα. Και ευχαριστώ για τη δουλειά του, πιο πιο ότι ήταν μια μονοσυκλέτα, αλλά όχι δεν είναι πιο λεπτά. Πυσικά, Μαρκούς πιο πιο πιο χρήκο, για τα σχέση της δημιουργίας της Ρετίνας, και όλα τα χρόνια της Ρετίνας, από my site, πλήθειτε να βρήκετε τον Περισταρμείς. Μαρκούς, η δημιουργία είναι όλους τους. Ευχαριστώ Γιος, θα πείτε το εμπορία μου τώρα. Ναι, είναι πολύ καλό. Τις. Λοιπόν, καλώς. Και ευχαρδώσουμε, Γιος και Τον, και όλοι που είναι εξεγραφήτητας σε όλες την Εμβιτέση, είναι πολύ ευκαιρή να είναι στον Ρετίνα. Είναι έναν άνθρωπο εξοδοσιασμό για να φτιάξουμε πρόσφερα, σιχεία να βρήκουμε πάνω όταν την προσπαθία είναι πολύ δύσκολο. Πώς έγινε το πρόσφερο για αυτή τη συμφωνία σας. Έχω εξεριασμένη εξερία με παιχνίδιο. Είχαμε να δοκιμάμαι ένα νέο πιο, μετά με δύο άλλους. Ήταν να δοκιμάμαι ένα νέο πιο, μετά με δύο άλλους. Αυτό είναι πολύ εύκολο, είναι ένα νέο φινάμενο από την αριθμότητα που είναι πραγματικό να δείξουμε. Επίσης, κάθε πράγμα για αυτό είναι περισσότερο από το σχέδιο σχεδόμενο που έχουμε με το σχέδιο Ρετίνα. Και τότε, οι τρεις ρευότητες αντιμετωπίζονται να εξηγήσουν πως, στο στιγμό τους, το υποσχέδιο της Ρετίνας θα υποσχέδει το σχέδιο σχεδόμενο, όχι μόνο πραγματικό, αλλά όχι μόνο πραγματικό. Βρίσκω, τότε, ότι υπάρχει, ούτε σε τα μάνα των διεθνών, ένα σχέδιο σχέδιο της πράγματος που δείξει. Και δεύτερο, ότι το σχέδιο σχέδιο είναι πραγματικό, να μην υποσχέδει τα παράδοια που ήταν βασικο, αλλά και να προσπαθεί ένα φινάμενο που δεν είχε been reported yet. Και είναι ένα σχέδιο που δείξει τις δυο πρόσφυρες από τα σύγγελια της Ρετίνας, για την αντιμετωπή του χρησιμοπού. Προσπαθεί με πολλές άλλες περιοχές της σχέδιας, νομίζω ότι η κατασκευή της Ρετίνας έτσι έτσι έτσι αυτοί και αυτοί της σχέδις. Και ο σχέδιος μου είναι να μιλήσω το σχέδιο, τι είναι το σχέδιο σχέδιο. Και θέλω να εξηγήσω, τι ήταν οι ευρωπαίες δυο πρόσφυρες που δημιουργούνται η κοινότητα για να έρθουν στον κόσμο της κατασκευή. Και θα εξηγήσω, λοιπόν, ότι θα έρθουν στον κόσμο της κατασκευή. Και στους τελείων, θα μπορούμε να συμβείξουμε πώς θα υποχωρήσουν κάποιες λεσσόνες για να κατασκευθούμε άλλες πράγματα της σχέδι. Τώρα, θέλω να κάνω ένα κομμάτι της κατασκευής. Πρώτα, αυτό είναι ένα επίπεδο της κοινωνικής. Είναι για τα εξογραφή από την κοινότητα της κοινότητας. Θα εξηγήσω με κάποιες δυο πρόσφυρες. Και, νομίζω, ότι θα υποχωρήσουν δυο πρόσφυρες για να εξηγήσουν δυο πρόσφυρες. Αν ή δεν υπάρχει κάποιες δυο πρόσφυρες, δεν υπάρχει αρκετά. Είναι πολύ πολύ σύστημα για την κοινότητα. Και, εύκολο, θα μπορώ να εξηγήσω μία πράγματα. Πρώτα, δεν θα πω να πω ότι όλοι έχουν κατασκευθεί το ρεθνό, ακόμα στον κομμάτι της κοινότητας. Υπάρχουν πολλές πρόσφυρές που εξηγήσουν. Αυτό που θα πω είναι πολύ σύστημα για την κοινότητα που εξηγήσει σε άλλες αρκετάς της κοινότητας. Υπάρχουν κάποιες δυο πρόσφυρες να υποχωρήσουν δυο πρόσφυρες, αλλά όμως, είναι πολύ σύστημα. Και, ευκολο, είναι το κοινότητα της κοινότητας. Ενώ, έχω πει ότι το σημίνο may be infiltrated by people who are not vision specialists. Αν αυτό είναι καλύτερο, είμαι very happy about it, but just to get everyone on the same page, the talk is going to be about early visual processing implemented by the retina in the eye. The rest of the eye serves as a camera using the cornea and the aperture of the pupil and the lens to project an image onto the back of the eye. That's where the retina resides, a layered neural tissue has the job of converting light into neural signals to then start filtering those signals from the photoreceptors through the various layers of neurons to ultimately the retinal gangon cells. The gangon cells are the output neurons of the retina. Each of these neurons has an axon, an optic nerve fiber, and the collection of those optic nerve fibers sends signals to the brain. The action potentials of retinal gangon cells are the only thing the brain knows about the visual world. Everything else downstream is based on those spike trains. Okay, so what is the standard model? Let me start with some basic premises. When I was in college, I learned quantum mechanics from a book that took a totally axiomatic approach. The book laid out the four axioms of quantum mechanics and then the rest of the book just derived all the phenomenology from those axioms. I think that's a fantastic ideal to aspire to. I'm not sure we're quite there in neuroscience, but nevertheless I tried to lay out what are some four premises that we believe and then we'll flesh those premises out in the form of what I think is the standard model of the retina. The first is there's a pretty clear conception of what the purpose is of the retina. Its job is to make light visible. The processing in the retina accounts for a lot of basic aspects of our visual perception and I'll give some examples in a moment. Given that what's the goal of the standard model? It's to explain and predict the responses of retinal gangon cells which are the output neurons of the retina as a function of visual patterns on the photoreceptors. If we can accomplish that to some extent the contribution that the retina makes to human vision. Another premise there seem to be about 30 or 40 types of retinal gangon cells so I'm not saying there's 30 or 40 of these neurons, there are 30 or 40 types of these neurons in the human retina there's a million of these neurons total. Each of these types is the output of a circuit that leads out away from the photoreceptors to the retinal gangon cell are distinct by exactly what that circuit looks like the parameters of the circuit model and I'll go into some detail about what that is. And finally retinal neuroscience has benefited a lot from comparison across species different species have retinas that look very much alike but they differ from each other in the types of responses that you find or that predominate in the retinal gangon cells and these types of responses again can be summarized by certain parameters of the circuit model and the thought is that different species have undergone a different selection for these parameters as a result of evolutionary pressure and ecological niches. Okay let me start by let's say this premise number one that retinal processing accounts for basic aspects of visual perception and just for background give you some reminder of some things that we know in this area. So very basic question how visible is light of different wavelengths and there is a long history to investigations of this problem and here is a result that was published a long time ago in yellow the data points are the visibility of light as a function of wavelength So in practice you have a human subject sitting in a dark room and you produce flashes of different wavelengths of light and you ask did you see it or did you not see it and you make the flash dimmer and dimmer until they stop seeing it and here is the sensitivity as a result and obviously it looks like we're very sensitive to light of 500 nanometers and somewhat less sensitive to light of 450 or to light of 550 and when things tail off beyond that. Now there's a second set of data points here the blue ones and these are biophysical measurement of the absorption spectrum of rhodopsin So you isolate rhodopsin from the retina put it in a cuvette shoot light through it of different wavelengths and ask how much light comes out the other end of the cuvette and this is the result in blue and obviously you can see that the points lie right on top of each other and more or less everything about the human spectral sensitivity is explained by the biophysical properties of one molecule rhodopsin So I think that stands as kind of an ideal of scientific reduction that one might want to aspire to So if there are other aspects of human vision that we can reduce to elementary molecular cellular or circuit properties to the level of precision of this example then we'd be very happy Okay let's get a little more complex How visible is light of different spatial patterns Again this experiment has a long history You show a human subject spatial patterns like the sinusoid gratings that can be either very coarse or very fine and then you ask again did you see it or did you not see it and you plot the visibility of the grating as a function of spatial frequency and that's shown in these curves In bright light you find that we're most sensitive to light of about five cycles per degree meaning there are five of these per degree of visual angle Just to remind you a degree of visual angle is roughly the size of your thumb when you hold it out So we're most sensitive to about five zebra sprites on your thumb and we are less sensitive to both coarser patterns which is interesting but in finer patterns In dim light the situation changes somewhat We are less and less sensitive of course to seeing patterns of light and it becomes more of a low pass kind of relationship meaning we're more sensitive to the coarser patterns Here's the equivalent measurement from a neuron in the early visual system of the macaque This is a neuron in the lateral geniculate nucleus of the thalamus but for all intents and purposes retinal gang cells behave the same way Again you can plot to the sensitivity of the neuron namely it changes its response as a function of the spatial frequency of the grating and again you find this kind of band pass relationship where there's a maximum around five cycles per degree and then the sensitivity trails off on both sides A good number of experiments of this type suggests that this basic aspect of human vision contrast sensitivity to different frequencies is essentially explained by retinal processing because we see the same relationship in the retinal gang itself You can ask the same question about temporal patterns so now the light flickers in time rather than being sinusoid in space it's sinusoid in time and you can ask as human subject can you see it or not see it and here's the sensitivity curve to the light of different flicker frequencies less sensitive to light that flickers at about 10 hertz and then less sensitive to very rapid flicker but interestingly also less sensitive to slower flicker so there's sort of an optimum of temporal variation that the humans are sensitive to and again you can do that same experiment with the output of the retina and you find a similar relationship band pass relationship for flicker frequencies and as in the spatial case if you go from bright light to dim light the relationship shifts to the left ok so these are some examples how aspects basic aspects of human vision can be explained by the output of the retina and that really motivates why one wants to understand how the retina works if we can explain how you go how nature goes from visual signals and photoreceptors to the output of retinal gang and cells then we've captured at least these basic elements of human vision ok so what are the ingredients of that standard model so we can take inspiration from other standard models in science then for example the standard model of particle physics which I understand at the level of popular science magazines but at that level the standard model specifies what are the components of matter what are the basic particles and what are the forces between those particles and the interactions and how can we given the particles and their interactions explain the emergence of bigger things like the nuclei of atoms or people ultimately so similarly here we're going to start with what are the particles what are the components of the retina and then what are their interactions and then how can we put the particles and the interactions together to explain something like circuit function ok the components in the case of a neural system are the cell types different types of neurons and in the retina there seem to be about 100 types of neuron I like to say it's a little bit like the complexity of a radio it's an old fashioned radio if you break it open and spill it out on the table you'll find about 100 different kinds of parts so traditionally the cell types were defined by their shape and here's a graphic representation of that for the receptors horizontal cells typically 2 or 3 types bipolar cells maybe a dozen types amicron cells maybe 30 types gangon cells 30 to 40 types and again the traditional distinction was by anatomy looking under the microscope at the states of these neurons they really have dramatically different shapes in many cases so that already serves to distinguish them from each other in recent years we've also been able to identify these cell types by their gene expression patterns either using distinct molecular markers or more recently through single cell RNA sequencing identifying more or less the entire gene expression profile and then these cells fall into clusters based on their transcriptome and these clusters it turns out can be identified more or less one to one with cell types that were previously defined based mostly on the shape now a number of you are probably wrinkling up their noses because I would say both anatomy and single cell RNA sequencing are very much forms of art in the sense that the classification that you arrive at is to some extent in the eye of the beholder anatomists are famous for seeing multiple types among the items that they inspect on the microscope but similarly single cell RNA seek is not a precise science and depending on how you set the parameters of the analysis you might find more or fewer clusters in these gene expression patterns so there's something remarkable about cell typing in the retina in that we actually have a separate criterion that tells us when we've identified a natural cell type maybe a type that nature believes is a cell type and this criterion has to do with the spatial arrangement of the neurons on the retina here's an example this is a face view of the retina with certain cell type labeled in red these are so called starburst amicron cells and you can see that these cell bodies of these neurons are spaced more or less regularly apart it's not the beautiful square lattice but there is a clear sense in which the cell bodies tend to avoid each other they keep at a certain distance from each other now you might not believe that from just staring at this one micrograph but if you analyze the distribution of these cell bodies relative to each other you can plot the probability of finding a neuron as a function of distance from another neuron of the same class and you find that there is this distinct hole at short distances meaning that there are far fewer neurons near another neuron than you expect by chance if we had sprinkled the cell bodies onto the retina like rain independently of each other this curve should be a constant the probability of finding another neuron as a function of distance should be constant instead there is this dip at short distances I expand on this a little bit because it I think is a very important benefit that we have enjoyed in studying the retina that these cell types are actually certified in this way if you look at the relative relationship of two different cell types let's say a starburst amocrine cell and a different type of amocrine cell they do not have this repulsion but that tells you that nature considers these two different types they are certified by biology as cell types the purpose of this of course underlying purpose is likely that retinal development spaces these neurons apart so that every point on the retina is close to one neuron of that type and you don't get accidental clustering and holes in the distribution you perfectly cover the visual space with all of the elements that are necessary for visual processing ok so much about cell types what about the connections between cell types so here again there is a lot of orders a lot of structure in how the retina is organized so at the courses level the retina consists of five layers there are three cellular layers the outer nuclear the photoreceptor layer a rental gangon cell layer and the inner nuclear layer and they are separated by two layers of synapses and these this alternation between cellular layers and synapse layers pretty much constrains the connectivity of the major types of neuron in the retina so photoreceptors connected by polar cells but not to gangon cells and by polar cells connected to gangon cells but not back to photoreceptors and so the five major classes of neurons are restricted in how they connect to each other just by the layering of their organization in the retina actually the layering is much more precise so there is this big synaptic layer called the inner plexiform layer it's about 40 or 50 microns in thickness and it is purely synaptos it's just dendrites and axons of the neurons connecting to each other and within that inner plexiform layer there's a lot of structure individual types of neurons will ramify their dendrites in just a very thin lamina of this inner plexiform layer to be precise these lamina can be only like one micron in thickness so there is they're probably on the order of 40 perhaps different lamina in this inner plexiform layer and neurons that ramify within the same lamina will be able to make synapses with each other but if they're in different lamina they're just not in the kind of proximity needed to make a synaptic connection so this inner plexiform layer in this lamina organization really specifies which neurons are allowed to connect to each other and which ones don't now within that constrained there's still a good number of possible partners that might form and which partners actually make synapses is determined by certain cell surface proteins that make two membrane stick to each other or not or determine whether a synapse gets formed or not and in recent years there's been a lot of insight into the nature of these cell surfers molecules they often come in families with a combinatorial diversity like the d-scams and the proto-caterins and the rules by which they adhere to each other will specify which neurons form a synapse and which ones don't and so for example we have a pretty good understanding for the molecular determinants of this particular circuit that leads from bipolar cells all the way to the direction selective ganglion cells and the molecules involved in forming some of these connections or constraining them at least have been identified to some degree already ok so we know something about cell types and something about the interactions between cell types and of course from that come neural circuits and the constraints on the connectivity really lead to constraints on what kinds of neural circuits you find in the retina but there's still a good amount of diversity that fits within those constraints on synaptic connectivity and so generally speaking each ganglion cell has above it a circuit that leads all the way from the folder receptors down to the ganglion cell neuron and these circuits have to satisfy the constraints on synaptic connectivity that I mentioned and here are a few of the motifs that we find in these circuits one common motif that of course you find also in the nervous system or more or less anywhere is pooling where multiple neurons pool their signals into one place like the multiple cone photoreceptors get pooled by a large field bipolar cell an interesting motif very early in the retina is pathway splitting so the cone signal immediately splits into the on bipolar and off bipolar are excited by light and the off bipolar are inhibited hyperpolarized by light this has interesting consequences really for the whole rest of retinal processing that the pathway splits very early on into pathways of opposite polarities in particular those opposite polarities can be processed independently and subsequently combined again in very interesting ways another common motif is lateral inhibition cone photoreceptors through horizontal cells can inhibit other cone photoreceptors bipolar cells through amachron cells can inhibit other bipolar cells there's also feed forward inhibition amachron cells in a feed forward fashion inhibit retinal ganglion cells these are all motifs that you find elsewhere in the nervous system as well a lot of interesting things happen I mentioned this a moment ago through the reconvergence of pathways so the early pathway splitting were a cone tox to the line or 10 different types of retina both bipolar cells ultimately those signals can reconverge at the ganglion cells and this leads to quite interesting computations I'll show some examples of that there are some key nonlinearities in this circuit and it's important to point out the nonlinearities because if everything were just a linear summation there wouldn't be much interest there's not much you can compute not much you can develop by just linearly combining signals from photoreceptors so the key nonlinearities I think are photoreceptors have a lot of gain control strong adaptation effects and photoreceptors you also see gain control at synapses for example the bipolar cell synapse has a good amount of short term synaptic plasticity important nonlinearity is rectification probably every neuron in the circuit has a nonlinear input output function to some extent a lot of interesting neural computations can be explained by the nonlinearity of the bipolar output synapse what that means is that the depolarization releases neurotransmitter and hyperpolarization doesn't do very much and so the 40 different types of retinal gangin cell are essentially characterized by 40 different kinds of circuits that lead to the output and you can parameterize those circuits in this standard model by specifying well what is the receptor type rods or cones feeding the bipolar cell what are the types of bipolar cells in that circuit what types of amortrin cells do they feed what's the degree of rectification and the type of cellular output synapse so these can be listed as parameters of a circuit model and then used to actually make predictions for how that circuit would work so let me show you some examples of that and first of all I want to give you a sense of how well the standard model works so I've argued that we have an understanding of the retina that really helps to explain or predict phenomena at the level where it becomes interesting for understanding human vision and so what is that that level of prediction really like so here's an example of trying to use a standard model to predict the output of the retina we used here movies from the real world and projected those movies onto the retina of a mouse this movie was taken from a little video camera on top of the mouse this movie was taken by simulating a mouse with a roller skate putting a video camera on the roller skate and running it through the grass to try to get the right perspective on the world anyway we project that movie onto the retina in the laboratory and record the spike trains of a retinal gang and sell them here our responses to six identical repeats of the same movie these are raster plots so every tick mark is an action potential and you can see that on subsequent repeats of the same movie this gang and sell does very much the same thing the firing rate goes up and down by a lot from 400 hertz to zero in a few milliseconds and so it's very strongly modulated by what's happening in this movie it's very reproducible the little gray shadow here is the standard error around the average response so there is a lot to be explained here similarly this other movie the same retinal gang and sell produces very strong variations and firing rates again with a high degree of reproducibility ok so how well can we do in actually explaining this output of the retina here's an example for this particular neuron we write a circuit model that takes assumes that light in the center of the receptive field is processed by a set of off bipolar cells light in the of the receptive field is processed again by off bipolar cells but then through a set of amicron cells the gang and sell pulls these two signals from the center of the surround and there's rectification of the bipolar cell output so it's it's a relatively simple circuit I'm sure it is oversimplified compared to what's actually there in nature but it suffices to make a good prediction so here in green is the variation and firing rate of the neuron is a function of time that I showed you a moment ago and in red is the prediction from the model after feeding the movie through the simple circuit and red line and green line are in pretty good correspondence I'll say in a moment how good the correspondence is but let me just make clear that this is not an accident that only applies to the particular type of gang and cell and the mouse here's a similar result from macaque on and off parasol cells so this is a major class of red and of gang and cells in the primate retina and again the data and the output of a simple filter model like that are plotted on top of each other and you can see that the correspondence is quite good in fact in both cases the correspondence explains about 80% of the explainable variance so explainable variance is the amount that the firing rate varies minus than always in that measurement so this is to me 80% is a good result but people definitely fall into optimists and pessimists in this regard and sometimes the same research group will publish a paper where they celebrate 80% and then another paper where they complain that it's only 80% and how are we going to understand the remaining 20% so we each can fall on different parts of that spectrum to me the glass is 80% full and I feel like this is a real accomplishment if we can understand the spike trains at the output of the retina to 80% accuracy under the kinds of conditions that the retina works with in real visual life now I know this is we're not as far along for all the types of retinal gang and cell and maybe even all the types of natural movies but there's a clear existence proof and kind of a path forward and it looks to me like the standard model is going to with sufficient elaboration lead to this kind of 80% understanding across the board of the retinal cell types ok so far I've talked about just predicting the firing rate of neurons but obviously retinal gang and cell spikes and there's been a proposal at least that the precise timing of the spikes is important in signaling things to the brain and it turns out you can with a small enhancement of the standard model you can predict these spike times the only thing you really have to do is replace the retinal gang and cell with an integrated fire model integrated fire model means that the input to the cell gets compared to a threshold when it crosses a threshold the neuron fires a spike and then resets the memory potential by a certain amount that decays exponentially so this integrated fire model coupled to the input from bipolar and haemocrine cells and others are quite realistic looking predictions of spike trains so here's an example from a few different types of retinal gang and cell of real spike trains on a visual stimulation on top and the simulated spike trains from the model with the correct parameter settings on the bottom and it becomes hard to tell which is which the senses that the model can really capture the output of the retina down to the timing of individual spikes and the variation in the number of spikes and so on so I think again the substrate is there for making predictions down to a quite precise level of resolution so let's talk about some more exotic functions of the retina that go beyond linear filtering so here's an interesting phenomenon that's probably also connected somehow to human vision which is a pattern adaptation if you expose the retina for some period of time to the vertical bars the gang and cells become less and less sensitive their activity declines by like a factor of 2 over 10 seconds or so and if you then switch to horizontal bars the gang and cell the same gang and cell will jump back up to a high firing rate and decline again over 10 seconds by a factor of 2 and so there is this sort of stimulus specific pattern adaptation which has a direct parallel in human psychophysics and we've wondered for some time how that can be explained in retinal circuits and the alternate theories have been proposed and it looks like the one that's winning out today or as of a few years ago is short term plasticity at the terminal of bipolar cells turns out that individual bipolar cells can become selective for vertical or horizontal bars because of asymmetric inhibition from an amicron cell and then if a bipolar cell is very active under that stimulus probably its vesicles get depleted short term synaptic depression sets in and the gang and cell gets less and less input over time whereas when the pattern switches to a different orientation a different set of terminals that is still at full strength now is active and it gradually depletes and becomes tired and leads to this pattern adaptation so the pattern adaptation used to be thought of as mysterious phenomenon I think fits squarely into the standard model here's another interesting phenomenon this is a bias for approaching motion this has been reported in the parasol cells of the macaque retina if you take a sort of a random texture picture and enlarge it gradually which is similar to what might happen if the animal approaches the texture these parasol cells fire strong bursts of spikes whereas in the opposite motion when the pattern recedes they they hardly respond at all so and this is true for both on cells and off cells and it doesn't matter what the exact structure is of the texture and this incidentally was the paper that we reviewed and argued that its predictions or its phenomena are predicted by the standard model and in fact if you just take a single ingredient of the standard model namely the integration over nonlinear subunits which are likely bipolar cells you can predict that the output the retinal gang and cells will respond more strongly to approaching motion than to receding motion and there's a whole bunch of other exotic computations that the retina performs that similarly can be explained with the standard model circuit the certain retinal gang and cells that are selected for the differential motion between an object and the background and the circuit has been proposed for how that might work and the circuit has been tested in fact by injecting current into these neurons and seeing that it has a proper effect on the output a different kind of looming sensitivity has been traced again to a version of the standard circuit using these amateur cells as interposed neurons between bipolar and gang cells I mentioned a moment ago that there have been proposals for how spike timing but encode visual information that's of interest to the brain and again the timing of specific spikes under different stimuli can predict that quite well, it can be predicted quite well by the standard model of neural circuits and other functions, more esoteric functions by which retinal gang and cells seem to switch between on and off type depending on context again can be traced to the circuitry that is sustained by the standard model of the retina so the sense is that we are really doing quite well in not only explaining phenomena that have been observed but even in trying to predict phenomena that we don't even know about yet that's actually a branch that hasn't been exploited enough I think I think it would be worth working through what are some possible predictions that haven't even been tested yet experimentally and that's I think would be a great rotation project for students in your laboratories ok let me move on a bit and summarize why do we think that retinal neuroscience has come to the state of affairs where the important aspects of neural function can be captured at least to like 80% accuracy so here are some ingredients of success that we've already talked about but there's a pretty clarity of purpose of the retina it's there to process light it doesn't have to do with hunger or memory so we can pretty clearly say that it's involved in these functions that have to do with processing visual patterns whereas if you're working somewhere deep in the brain on the hippocampus there is no clarity of purpose or what that structure does for the animal there's a really close link to psychophysics we've talked about it already but for literally hundreds of years people have experimented on their own visual system and there are some fantastic visual illusions that illustrate some of the oddities of vision that you can replicate them in the retina so this close connection to visual psychophysics I think has made a big difference especially in guidance like what problem do you actually want to understand we talked about cell types in particular cell types that are certified by nature that's important we talked about the structure connectivity that comes from the beautiful layering of cells and synapses in the retina experimental access this is an important one you can take the retina out of the eye put it in a dish and it works it may not work exactly the way it does in the eye but it works so a lot of the basic circuit functions are there and you can study them and you can then go back and verify them in vivo and make sure that this also happens in the eye of course taking the retina out into the dish gives you fantastic access with sharp electrodes and electrode arrays and pharmacology and calcium imaging and what not so the the access has been fantastic another thing I'd like to say is that the retina is naturally optogenetic many of our colleagues are struggling to make neurons respond to light so that they can inject signals into their neural circuit of choice the cone photoreceptors come with the option built in already and so you can stimulate 100,000 of the input neurons with pixels of light on your monitor and the this sort of level of resolution of stimulating the input to a network is really hard to attain otherwise and finally I touched on this but I think it's very important to solve the cross species integration of the discipline you know and 100 years ago Lord Arian started recording from eel eyes and you know already reported some of the phenomena we are still interested in today subsequently you know cats and monkeys and frogs and toads turtles and salamanders and more recently it's been popular to focus on the mouse retina but it's important that a lot of this cross species work has shown how reproducible the basic principles are the cross species and that has really been allowed us to focus on the important aspects of retinal processing I should say by cross species I would include invertebrates it's kind of remarkable how similar some of the retinal processing is in flies and mammals and this is just one example of the review paper that illustrates this for the case of motion computation and both flies and mammals you find four classes of neurons that respond to motion along the cardinal axes and the way these things are computed is actually quite similar in the two cases now these are hundreds of millions of years of evolution apart they obviously develop independently of each other and it's interesting to see how nature struck on a very similar algorithm for computing directional motion information about the image okay I also want to point to some things that were not necessary for success and understanding the retina and oscillations so very popular meme in brain science that neurons communicate with each other or brain errors communicate by oscillating I remember when I was a young assistant professor I visited a senior person at another institution who proceeded to explain that obviously the retina is just a series of coupled oscillators I mean just look at it right the neurons are organized in the plane and each neuron is an oscillator and you just have to figure out how they're coupled to each other and that will explain the retina so this was very much an engineering mindset that visual images are translation invariant the eigenfunctions of the translation operator are sinusoids and therefore obviously visual system should perform Fourier transforms and oscillate one thing we can say quite certainly I think 30 years later is that that's not the case there's no sense in which the retina performs a Fourier transform there's no sense in which it operates or signals as a set of coupled oscillators now there was a brief scare in the 1990s when a number of high profile articles reported that neurons in the visual cortex of the cat were oscillating at precise frequencies and in fact different neurons were oscillating at different frequencies and the claim was that this solves the binding problem that different parts of an object were identified by the same oscillation frequency and the different object would oscillate at a different frequency and that's how downstream parts of the brain know that the parts of the object belong together because they're all oscillating at 91 hertz and then subsequent high profile papers trace the oscillation into the lateral geniculate nucleus and eventually to the retina and in this report here there were electrodes simultaneously in the retina, the thalamus and the cortex and they all oscillated at 91 hertz and with incredible precision so at that point the retina community so raised its head and said well look we've had literally thousands of reports of responses of retinal gang and cells and a tiny minority of them has reported any kind of oscillation and none of them has reported perfect synchrony at 91 hertz so what's going on here and I think in the meantime we now acknowledge that these oscillations were an artifact of anesthesia and we just don't anesthetize neural circuits anymore when trying to understand how they work and it's remarkable to me actually how robust the retina is to oscillations the nature seems to go out of its way to stop it from oscillating you might think that a circuit that has a large gain like photoreceptors with gains of a million or so and feedback loops a circuit like that would be prone to oscillating but actually it's hard to make the retina ring at all you can do it with incredibly strong flashes of light you can get a few oscillation cycles but the system is designed to be over damped so that it doesn't oscillate another ingredient that I would argue was not necessary for understanding the retina at least so far is sort of a full blown compartmental modeling of neurons now you've all seen models like this where the neuron gets divided into 100 or even a thousand compartments and each compartment has its own configurations with separate memory conductances and they're coupled to each other and then you have to use a super computer to understand even how the single neuron works the model will have hundreds of parameters many of the parameters are unknowable or at least unknown to science at the moment and so it gets very complex at the single neuron level I would argue that we haven't needed that level of complexity but most of the neurons in the models that people draw and calculate through are point neurons meaning they integrate their synaptic inputs and they produce a synaptic output as a result with no local nonlinearities there are a few cases where we need to divide the neuron into two or more compartments so for example the bipolar cell soma and the bipolar cell terminal have different voltages because the terminal gets input from amicron cells for example and that inhibition does not travel all the way back up the axon to the bipolar cell soma so the bipolar cell needs to be divided into a couple or more compartments similarly the starburst amicron cell it's a very cool neuron that has a lot of processing happening within its dendrites in particular the dendrites are sensitive to object motion that goes outward rather than inward and each dendrite seems to be doing that computation independently of the others so the starburst amicron cell is another case where it's reasonable to divide the neuron into four or five compartments and consider them separately but I would say that the degree of single neuron biophysics that has been necessary at least so far is relatively modest compared to the full bone compartmental modeling that you see in some other studies in brain science here's another theme that is very popular in neuroscience these days low dimensional dynamics the idea is why does the brain have the millions and billions of neurons well maybe it's all organized in a very inefficient way so that the million neurons are encoding just two or three important variables and so projecting the activity of all the neurons that we record onto principal components one can then draw these trajectories in the low dimensional space to try and summarize the overall neural activity in the circuit and but people go beyond just using this as data analysis tool to proclaiming that it is a principle that brain dynamics is in principle low dimensional it happens on low dimensional manifolds like toruses and the fact that there are millions of neurons involved is just a complication that we have to try to suppress and see through this is not the case in the retina very simple argument the human retina has a million neurons as outputs retinal processing is organized so that those million neurons are doing more or less different things from each other they try to avoid being redundant with each other and so in some sense the retinal output has a million dimensions and they're irreducible you can't really claim that there is a lower dimensional manifold hidden inside those million dimensions let me skip this in the interest of time and talk instead about sort of output for the future so what should we what should we do with the standard model and where to go from here I think an important thing in the future is to acknowledge the standard model and use it as a reference so this can be used to plan new projects you could for example test the basic assumptions of the model one basic assumption is that the retina gets almost no feedback from the brain which allows us to think about it as a feed forward system but is it really true that there are some recent measurements that suggest that for example the state of arousal of the animal can have an influence on retinal processing I think that's very well worth following up there are a few optic nerve fibers that go in the opposite direction and so what do those do in order to modulate the function of the retina just one example you could take the standard model and ask which of the components do we know little about and for example there are these wide field amicron cells that send their axons clear across the entire eye no why fundamentally does this side of the eye have to know what the other side of the eye is seeing I'm not sure to me that's still a mystery but you know there are definitely components cellular components of the standard model that need to be elaborated another thing I think would be useful that we should do more of is to be quantitative about the contribution so if you report the new effect and that effect makes a factor of two difference in how the gang and cells work on certain conditions that's remarkable that's going to be important that's going to affect vision and so on five percent difference that's sort of less impressive I feel like it's useful to make reference to the existing state of knowledge and ask how much are you modifying it that's sort of on the experimental side I think on the theory side there is a more serious open more serious gap and I think we need a better theory for why the standard model is what it is and you know what kind of question is this so in particle physics you can ask why are the particles what they are and very soon the discussion goes into philosophy people say oh it's a result of the big bang and the future will have a big crunch and another big bang and it will be a different set of particles who will be in a different universe or maybe there are already a million parallel universes where the standard model is different in biology you don't have to become philosophically immediately because there is a sense in which this question makes sense and that has to do with evolution so the retina is the way it is because of the outcome of many evolutionary accidents genetic variation mutations have varied the parameters of the standard model and other forces like genetic drift or natural selection have led to fixation of some of the parameters and that's why it is what it is but we need to develop that into a real explanation and some of you might say well we have such a theory it's efficient coding theory the theory says that the retina tries to pack as much information as possible into as few spikes as possible and that explains the processing that happens in the layers of the neurons that lead up to the retinal gang itself and in fact this theory has been very well developed kind of blossomed in the 1990s and has been followed quite seriously ever since it makes good predictions for example early on it was clear that it would predict the center surround structure of retinal gang and cell receptive fields and the fact that light in the center excites the neuron and light further away inhibits it similarly it can predict the human kind of sensitivity curves we've looked at these data before sensitivity of humans to life of different spatial frequencies and some people argue that many other aspects of early psychophysics can be predicted by the same efficient coding theory but I would say that that really doesn't explain very much about the retina the main predictions of the theory are center surround receptive fields and the biphasic time course to the temporal sensitivity of the neurons well these properties are already developed in the outer retina bipolar cells have these kinds of receptive fields and so the question that comes up why do we need 40 different types of gang and cells each of which has the center surround receptive field in order to encode the output from the retina with a little bit of arm twisting you can make the theory explain why there are on and off pathways in the retina although I'm not totally convinced that it's a very satisfying explanation even though I tried it myself but we really don't understand why there are 40 types of gang and cells but we don't realize them in these particular ways and report the information along 40 different pathways to the brain so I think there is a lot of room there for a theory that does better and in looking for a new theory I feel we should take a bigger viewpoint and ask you know what's the purpose not just of the retina but of the entire visual system and I would argue that the purpose of the entire visual system is to find a needle in a haystack Roughly speaking the eye gets visual information at a rate of about a gigabit per second here's a little back of the calculation to illustrate that and then the brain that's attached to the eye extracts from that about 10 bits per second so here's a human typist that converts a scroll of a pencil on yellow paper into a manuscript and she produces letters at an information rate of 10 bits per second it's generally true that the throughput of human behavior is about 10 bits per second no matter what we do but of course we can change what we do from one moment to the next so the visual system that supports these visually driven behaviors has to find these precious 10 bits per second in this giant junkie or the 9 bits per second I think that's the fundamental challenge of the visual system as a whole the challenge is one of computation it's not one of efficient coding efficient coding theory says you want to maintain the information and just package it differently that's not what the visual system is trying to do is trying to find just a tiny number of bits and throw away all the rest and so we need to think more about how does the retina throw information away how does it preserve the information and sort of back of the interval calculation suggests that in fact the retina passes through only maybe 5% of the information that's present in the corn folder receptors and discards 95% that gets turned into heat so I feel like a better guide to a theory of the retina is going to be focusing on computation rather than coding and but this is very much an open area I think that we need to make progress in this field anyway that gets me to the end let me thank some of the people involved my current research students in my current research group some recent graduates are responsible for some of the work that I reviewed and then recent collaborators that are also steering me into new directions in brain science thank you very much thank you very much marcus for these amazing talks celebrating the success in an overview manner of the retinal neuroscience community and relating it to other fields of brain science if I was about to ask you if you could stop saying just so we appear bigger on people's screen so given that it's one of the rare opportunities maybe not even rare a single opportunity in the two and a half years that we have been having this Sussex Vision series we mostly focus on vision or retina or invertebrate and today we have with us hopefully a much broader audience coming from AI machine learning, studying cognition dendrites, manifolds and so on it would be amazing if many of you join for this post talk cheat chat that is still broadcasted so you can follow the link that I just posted but at some point we will stop broadcasting and continue in the zoom room exclusively there are already a number of questions that appear in the chat but I will start with one that appeared towards the end and generated some hype already so it comes from Will Kearney how well does the standard model as described here largely feed forward hold up to predictive processing accounts of perception ok so this is a question about predictive coding which has different meanings to different people multiple meanings anyway one way it is used is that the given stage of the visual system tries to represent the differences between its input and the prediction that is made by a higher stage of the visual system so for example if some higher stage of the visual system thinks that there is a face in a particular part of the visual field that would be projected down to the earlier stage of the visual system and make it process that part of the visual field differently so obviously that information that kind of high level interpretation is unlikely to come back to the retina as I mentioned there are a few optic nerve fibers that go from the brain to the eye but not enough to deliver sort of high resolution or high dimensional information about what's expected in the particular part of the field on the other hand the retina itself performs predictive coding in its own its own primitive versions for example lateral inhibition and in particular feedback inhibition is a form of predictive coding they surround of the ganglicell receptive field makes a prediction for what the intensity ought to be in the center and that prediction gets subtracted from the actual intensity and the ganglicell reports the difference fundamentally the ganglicells are differential encoders they report the difference between the predicted visual scene and the actual visual scene using predictions derived from a simple model of the world like the world is made of objects and objects have some visual extent and therefore I can predict the intensity in this pixel from looking at the surrounding pixels and similarly in time I can predict the intensity now from looking at the intensity earlier because most of the time it doesn't change so that kind of predictive coding happens within the retina I doubt that the retina is tied into any kind of hierarchical predictive coding structure right, thank you very much for addressing this first question and what I inappropriately didn't manage to convey in time that you have a lot of messages both greeting you at the beginning you for the talk at the end and all these messages will be appearing in the youtube so in case you want to see them later next question again going from a general to more specific perspective is from Matthew Yadutenko have you tried to test standard model in a framework of similarity preserving normative theory I think the short answer is no I'm not entirely sure I understand what theory is being referred to Matthew is here with us in the zoom room already so Matthew in case you would like to clarify that would be great if not I can proceed for the time being with the next questions that appear ok, so then I will move to the next questions for the time being it's one from Jonathan Royce it's a non-specialist question and it was like when you were showing the retina overview with the electronic circuits the bipolar cell non-linearities and the question is like what kind of behavior I guess he means like electrotonically speaking does plasticity and rectification mean in a bipolar sign-ups ok, good point ok, so rectification refers to the instantaneous input-output function so output of the neuron and it typically has a kind of a threshold you know the machine learning people would call the relu relationship where below a threshold there is very little output and above the threshold it increases linearly for some neurons we also have to include saturation because they can only increase their output up to a certain level and if stimuli or inputs are very strong then you get addition to the threshold behavior you also get saturation at the top so that's what I mean by rectification it's the instantaneous relationship between input and output plasticity or adaptation refers to a change in that relationship over time so for example if we let's take the relu simplification where the non-linearity is just a threshold followed by linear function if the neuron has been very active for some time the gain of that non-linearity will decrease so the output becomes less sensitive to the input if the neuron is silent for some time it recovers and the gain increases again so that's what's meant by plasticity or adaptation is a time dependent and an activity dependent change in the gain of the relationship the rectification refers to the relationship itself having this kind of threshold non-linearity and staying on the computational aspects of your talk the next question is from Madinesa Vestani why does X not being relevant for explaining and coding variance in the retina mean it's not relevant for the rest of the visual system with its recurrence, multimodality and six-layered structure ok i didn't mean to imply that we cannot take the retina as extending to the rest of the brain so if i say that for example full bone compartmental modeling was not necessary to understand the retina that doesn't mean that it doesn't play a role anywhere else it does mean that there's an existence proof that you can explain fairly complex phenomena without it and so it might encourage people working on the cerebellum or the insect mushroom body or something like that to wonder you know at least try to see what can be done without assuming that the individual neurons are hyper complex computing engines can we actually capture the phenomena with a simpler model of single neurons it it does not exclude that more complex models of single neurons are necessary in other cases but i feel like there is a tendency to think that the brain is the most complex object in the universe all these fantastic functions of cognition it has to trace down to the complexity of single neurons and i feel like our experience in the retina has been that even with simple single neuron models and not a terrible amount of complexity of circuitry you can explain a lot of phenomena that seemed puzzling to begin with thank you very much for that before i continue with the next questions that appear i have one of my own so you mentioned retinal neuroscience success as a community that we are doing at least 80% of explained variants and this would be celebrated my question goes now to vision restoration and there like we have many attempts like with electronic with electronic implants or biological or introducing some options and so on but i guess we are not at the same stage of success and i know that this is relative like for someone that is blind even being able to see some shades it's already a huge huge success but why do you think this field is still like compared to the physiology of retina it's still at its infancy like is it because we don't have a mature understanding is it because we lack the equipment to restore vision or somewhere in between i think it's not a lack of understanding the retina and it's everything to do with getting access to neural signals that get sent to the brain i mean when i started in this field people began to implant electrode arrays and eyes with the idea that after the photoceptors have degenerated you can then stimulate the gang and cells directly and send pulses through the optic nerve to the brain and make the brain give the brain the illusion that the eye is still working and i don't know to me it always seemed like a far fetched idea technologically speaking just because you know you put the grain of sand in the eye the retina shrills around it and how do you keep that from happening and you know we're not pretty sure that that's not gonna work you know a company was founded around that idea several companies have been founded around that idea have actually implanted electrode arrays in the eyes of patients patients never emerged from blindness and now the company's gone bankrupt and people are stuck with these things in their eyes and nobody there to service them so it's kind of an example of technology failure i think there are much more promising avenues to vision restoration but they're not they're not implemented yet right so one obviously is you know stem cell therapy of some kind getting rods to reform and integrate into the circuit after or you know ideally before the generation happens completely i think the optogenetic approach is very cool you know putting transducers into rental gang and cells or even earlier neurons like bipolar cells and driving them directly with light after the photoreceptors have died i think there are clinical studies ongoing and i think that's a very promising direction in our own lab we've tried a different approach so okay so let's back up for a moment the challenge is that had refers to my last slide the retina gets about 10 to 9 bits per second of information going into it into the photoreceptors and the human being ultimately needs 10 bits of that so so these retinal approaches to vision restoration are trying to feed information into the eye the rate of 10 to 9 bits per second they're trying to reconstitute the flow of information at the input and that just seems very hard to me it seems very hard to build a brain machine interface that can operate at that kind of bandwidth and you might say okay the 1% of that might be good enough but it's still a lot of bandwidth that is needed and fundamentally the retinal implant company failed because of the lack of bandwidth so why not take the opposite approach which is to give the person the 10 bits per second they're looking for and so for example for reading that problem has been solved if text-to-speech devices that will read what's on the screen for you and you can get that information quickly as if you were reading it with your eyes and so our thought was can you read the rest of the world as well can we turn the visual scene into audio representation that tells the viewer so to speak what is out there without requiring the visual parts of the brain and so personally I think that's an avenue that ought to be pursued more we should communicate with the brain at the 10 bits per second that it needs in order to function not at the 1 gigabit per second and then hoping that the visual system will correctly sift that out again so that it gets only the 10 bits per second that it needs and I feel like the kind of assistive technology that exists now fantastic tools for automatic image recognition and interpreting the contents of images and annotating in real time 60 times a second we need to make use of all the clever machine learning and AI tools that have been developed in that area and make prostheses that work at the cognitive level rather than trying to reconstitute the basic neural signals of the retina the encoding itself thank you very much for this detailed response Marcus and I will be taking one last question that appears in the chat before stopping the live broadcast so I would like to thank our audience as well for being here and ask them if they want to be part like as an audience or actively participate in the ongoing conversation to make sure to follow the link I will repost right now so the last question I will be taking from the chat is from Philip Bartel generally what aspects of the message change when the activity of the neuron changes is the form of the message change somehow or is it always only the value that changes I don't know if this is hinting at the latency and coding or if it's just like this distinction okay, alright um um um um um um um um alright um so I think the question might get at what is it about the gang themselves spike trains that actually conveys information to the brain and this is an old question and not only in the case of the retina but in general people have been contemplating what is the neural code what is it about that is important to downstream regions of the brain um and uh there are two extreme views of opinions on this one is that the only thing that really matters is the number of spikes that the neuron fires over some period of time let's say a tenth of a second and so we can summarize the message that the neuron like a gang itself conveys by just listing the firing rate averaged over let's say a tenth of a second um the other extreme view is that no every spike is sacred every spike has its own message to convey and it's a precise timing of those spikes that is important for downstream visual processing I feel like in the visual system we probably have a better understanding of these at least the constraints on these theories than in some other parts so for example uh some aspects of visual perception are incredibly fast there are decisions you can make you know maybe not consciously but nonetheless that you make within a tenth of a second of the visual input or two tenths of a second and uh if you ask how does the visual signal proceed through different stages of the brain you know starting in four receptors several synapses to the original gang and cells several more synapses through the visual cortex it's clear that uh the neurons along the way cannot be counting many spikes uh they have to operate on either single spikes or just a few spikes there is no time for the system to calculate the firing rate of the neurons and um so as a result there's been a lot of interest I think in taking single spikes seriously in the output of the retina and also of course in subsequent stages of the visual system because they are probably involved in at least some very early steps of visual processing that might be performed just based on one spike per neuron thank you very much once again Marcus for addressing all these questions and thank you very much to the audience that was here with us today I will be stopping the live broadcast right now so if anyone was sigh until now and already in the room you can start your video because we are going offline thank you very much