Characterizing nonlinear auditory computations

Johns Hopkins University, Baltimore MD

Investigators

Abstract

Many neurons in the auditory system respond to sounds nonlinearly; that is, its response to two sounds played simultaneously differs from the sum of its responses to each sound played alone. Nonlinearities are necessary for many computational functions, but unlike nonlinear models that allow closed-form solutions, nonlinear models are often too hard to characterize in practice. To make nonlinear models tractable, this project will combine single-unit recording in awake marmoset monkey with automated online stimulus design by parallel computing. The goal of this stimulus design is not to maximize the firing rate of a neuron, but to extract the most information about the global stimulus-response relationship. Optimal sounds will be designed "on the fly" according to a neuron's response history, with the help of a fast parallel computer whose running time is compatible with the single-unit recording experiment. The proposed research is expected to produce practical and widely applicable methods for characterizing nonlinear sensory neurons. The auditory system is an ideal system for this type of online experiment because sound space is of lower dimensions and allows faster computations. The methods developed here are expected to generalize to nonlinear problems in other sensory modalities. Theory and algorithm development will focus on generating sound stimuli which can either most accurately estimate a given model, or maximally distinguish competing models. Nonlinear models with various degrees of complexity, including neural network models, will be used simultaneously, and contrasted against one another in the automated experiment. The model-based sound design method will be used to characterize complex response properties of neurons in auditory cortex and inferior colliculus of awake marmoset monkey, a vocal primate. This project focuses on the auditory cortex because studies of its pronounced nonlinearities may potentially benefit most from the new method. For comparison the same method will also be applied to the inferior colliculus, the inputs to which are better known, allowing more realistic hierarchical models to be developed. The models obtained from this method should provide a concise summary of the global stimulus-response relationship of a neuron that generalizes across all types of stimuli. Neural network models may also help extract additional information about the connectivity between different neuronal types, thus providing a link between the stimulus-response function and the structure of the underlying neural circuits.