Stanford OnlineStanford EE292R AI and Data for Science | Autumn 2026 | Models of Computation in the Retina
Stanford Online
Retina as a model 0:08
The talk starts by using the retina as a way to study how neural circuits compute. The retina is simple enough to measure, yet rich enough to show how cells and synapses build useful operations. The goal is to connect what the retina does with the machinery that makes it happen, and then to use that same approach for other neural networks.
Model matches retina cells 21:00
The model’s interneurons matched real retina interneurons under identical stimuli, even though the model had no prior knowledge of those responses. The strongest matches were close to the natural correlation between real interneurons, around 0.6, while model-to-cell matches were a little above 0.5. That was enough to turn the model from a predictor into a source of real insight and testable hypotheses.
Natural stimuli reveal functions 23:01
Natural scenes let the model reproduce many known retina effects without being trained on them directly. It adapted to contrast, changed response timing with stimulus strength, showed frequency doubling, polarity reversal, predictive firing, and omitted-stimulus responses. It also reproduced object motion sensitivity when the training set contained natural movies rather than jittered images. That same approach was then used to compare retina and cortex and to ask what each cell was doing inside the circuit.
Reading out causes 33:31
To understand computation in a cell, the talk treats its effect as a map from receptive field to output field. In a nonlinear system, simple activations are not enough, because the answer depends on the path you take through stimulus space. Integrated gradients solves this by summing each cell’s contribution along a straight path from a blank stimulus to the real one. That lets you see which cells actually caused a response, not just which ones were active.
Path analysis and probes 43:00
The discussion turns to path-based analysis in the retina and in simple vision networks. Looking at whole paths, not single neurons or synapses, makes it possible to measure how excitation and inhibition cancel each other. In a simple digit-recognition network, that cancellation was about 46 times after ordinary training, and only dropped to about 18 times before the network began to fail. The same method was then used to track changing receptive fields in the retina and to see which paths drove them.
What the methods revealed 47:01
These tools showed that receptive fields are dynamic and can shift with the stimulus. The speaker says the work suggests vision models can be accurate, generalize well, and still be explained mechanistically. The same ideas may help uncover new phenomena and generate testable hypotheses about both the model and the real biological system. Questions from the audience then focus on stimulus choice, color, eye tracking, and moving from salamander to mouse data, with the answer that the retina across vertebrates is similar enough that the structure did not need major change.
Sharing data and models 58:30
The final exchange points students to the published datasets and trained models, which are linked in the papers. The advice is to choose a project that is tractable and to use one of the machine-explanation methods already mentioned. The session ends with a note that a later talk will use a different kind of cortical model, not one that matches individual cells one by one.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
