UC Berkeley

The human genome (containing around 1E10 bits of information) is unlikely to fully specify the connectivity between neurons in our brains -- such a ``wiring diagram" requires around 1E14 bits. Physiological evidence suggests that the genome instead specifies plasticity rules through which the brain self-organizes in response to experience. As systems neuroscientists, we seek to understand those rules and, by extension, our brains. In this thesis, I will use this approach to study the primary visual cortex (V1) -- the brain region that receives visual inputs from the eyes, via a relay station called the lateral geniculate nucleus. I first study the statistical structure of natural images, which provide the visual experience that shapes V1. Then, I introduce a biophysically motivated model for visual cortex, which adapts to natural image statistics in order to efficiently encode them -- in this case, the neural plasticity rules can be shown to optimize this ``efficient" representation. I then demonstrate that this model can account for several features of V1 physiology, including the features to which V1 neurons respond (``receptive fields"), and the developmental trends in the sparseness of V1 activity. I will conclude that efficient coding models can be implemented within the constraints imposed by the neural substrate, and that efficient coding principles may yield a parsimonious systems-level understanding of visual cortex.