Dec310:00am - 11:30amZoom
“Geometry of Object Representation in Visual Hierarchies”
Haim Sompolinsky, Ph.D.
The Hebrew University
Abstract: Neurons in object representations in top stages of the visual hierarchy exhibit high selectivity to object identity as well as to identity-preserving variables, including location, orientation and scale. suggesting that changes in the object representations from low to high processing stages are related to changes in the geometry of object manifolds. Each manifold consists of the set of population responses to stimuli belonging to the same object.
In my talk, I will present recent work that elucidates the relation between manifold geometry and object-identity computations. I will discuss two kinds of computations. The first is object classification. I will describe new measures of manifold radius and dimensions that predict the ability to support object classification (Chung et al., PRX, 2018). Based on these measures, we characterize the changes in manifold geometry as signals propagate across layers of Deep Convolutional Neural Networks (DCNNs). Recordings from neurons in various stages of the visual systems, have been similarly analyzed, allowing us to test the correspondence between DCNNs and the visual hierarchy in the visual cortex.
In a recent unpublished work with Ben Sorscher (Stanford), we have studied the ability to learn new objects and object categories from just a few examples (the few shot learning problem). We show that feature layers in DCNNs exhibit a remarkable ability in few shot learning of new categories. To explain this performance, we develop a new theory of the geometry of concept formation, that delineates the salient geometric features that underlie rapid concept formation in artificial and brain sensory hierarchies.