UC San Diego

The sensory nervous system converts external stimuli into electrical signals that are used to process and transmit information about the stimuli. An ongoing goal of systems neuroscience is to describe the processing of stimuli as compactly as possible using a small number of features from the stimulus, which is known as dimensionality reduction. This task is especially difficult when analyzing stimuli with complex correlations between dimensions as is found in natural stimuli. This dissertation begins by presenting Maximally Informative Dimensions (MID), which selects the features that modulate the neuron's responses the stimuli. It then presents three variants of this method that seek to address specific limitations of the method. Sequential Maximally Informative Dimensions seeks to perform this analysis without calculating multidimensional probability distributions. Invariant Maximally Informative Dimensions allows simplified analysis of neurons that respond to similar but offset features. Quadratic Maximally Informative Dimensions incorporates quadratic features to allow one to find many linear features. Finally, Invariant Logistical Subunits combines the ideas of Invariant Maximally Informative Dimensions and Quadratic Maximally Informative Dimensions in a more flexible manner