Event:

How does the brain convert ever-changing, two-dimensional visual inputs into the seamless perception of a stable, three-dimensional world? This fundamental question lies at the heart of visual neuroscience and holds critical implications for artificial intelligence. A major cause of visual change is an animal’s own movement which transforms the images captured on its retina in a geometric way. I hypothesize that geometric image transformations shape how populations of neurons respond to rotated images, leading some neuronal subspaces to be rotation-tuned and others rotation-invariant. Analyzing mouse visual cortex recordings and convolutional neural networks revealed that early layers exhibit both rotation-tuning and -invariance, while late layers are primarily invariant. This work also highlighted qualitative differences between biological and artificial neural networks. To better understand the neuronal architecture underlying biological vision, I turned to the fruit fly’s visual system - a more tractable model system where all neurons and connections have been mapped. I developed computational tools to capture the spatial organization of neurons, which I then used to classify neurons into cell types and predict some of their visual response properties. Together, this multi-scale approach, which combines insights from mice, flies and AI, advances our understanding of how neuronal circuits transform raw visual inputs into robust representations of the world.