RT Journal Article
SR Electronic
T1 Sparse coding and dimensionality reduction in cortex
JF bioRxiv
FD Cold Spring Harbor Laboratory
SP 149880
DO 10.1101/149880
A1 Michael Beyeler
A1 Emily Rounds
A1 Kristofor D. Carlson
A1 Nikil Dutt
A1 Jeffrey L. Krichmar
YR 2017
UL http://biorxiv.org/content/early/2017/06/14/149880.abstract
AB Supported by recent computational studies, sparse coding and dimensionality reduction are emerging as a ubiquitous coding strategy across brain regions and modalities, allowing neurons to achieve nonnegative sparse coding (NSC) by efficiently encoding high-dimensional stimulus spaces using a sparse and parts-based population code. Reducing the dimensionality of complex, multimodal sensory streams is critically important for metabolically constrained brain areas to represent the world. In this article, we provide an overview of NSC, summarize evidence for its role in neural computation in disparate regions of the brain, ranging from visual processing to spatial navigation, and speculate that specific forms of synaptic plasticity and homeostatic modulation may underlie its implementation. We suggest that NSC may be an organizing principle in the nervous system.Basis functionsA lower-dimensional set of linearly independent elements that can represent a high-dimensional input space given a weighted sum of these elements, where the weight of each element is defined by a separate hidden component.Dimensionality reductionThe process of reducing the dimensionality of a space to the lowest possible space that encapsulates the variance of the original data via feature extraction. In the case of neuronal firing rate patterns, this means representing all possible firing rate patterns in the brain region using the smallest possible subset of the neurons.Factor analysisA statistical method used to describe variability among observed, correlated variables in terms of a potentially lower number of unobserved, uncorrelated variables called factors (or latent variables).Receptive fieldThe structure and boundaries of an individual neuron’s pattern of response to various kinds of incoming stimuli.Spike-timing dependent plasticityA Hebbian-inspired learning rule in which weight updates are computed based on the precise spike times of pre- and post-synaptic neurons that induce either long term potentiation or long term depression in the synapse, depending on whether the total pre-synaptic spike count preceded the total post-synaptic spike count, integrated over a critical temporal window.