CDS Thesis Seminar: Explicit Object Representation By Sparse Neural Codes

Stephen Waydo

Neurons have been identified in the human medial temporal lobe (MTL) that display a strong selectivity for only a few stimuli (such as familiar individuals or landmark buildings) out of perhaps 100 presented to the test subject. While highly selective for a particular object or category, these cells are remarkably insensitive to different presentations (i.e. different poses and views) of their preferred stimulus. This invariant, sparse, and explicit representation of the world may be crucial to the transformation of complex visual stimuli into more abstract memories. In this talk I will first discuss the issue of how best to quantify sparseness, particularly in very sparse systems where biases are significant, and show the results of this analysis applied to human MTL data. From there I move into the computational realm. Sparse coding as a computational constraint applied to the representation of natural images has been shown to produce receptive fields strikingly similar to those observed in mammalian primary visual cortex. I apply sparse coding as a model for processing further along the visual hierarchy:
not directly to images but rather to an invariant feature-based representation of images analogous to that found in inferotemporal cortex. This combination of sparseness and invariance naturally leads to explicit category representation. That is, by exposing the model to different images drawn from different categories, units develop that respond selectively to different categories. After extending an existing model of sparse coding and providing some mathematical analysis of its operation, I will show results obtained by applying this method both to unsupervised category discovery in images and differentiation between images of different individuals.