How neuroscience can benefit from advances in machine learning? the example of high-resolution retinal prosthesis.

Abstract:

Recent advances in experimental techniques have enabled large-scale recordings and stimulation of neural populations (e.g. multi-electro arrays, optogenetics) comprising thousands of neurons or even entire brains. The central question now is how do we make sense of these huge volumes of data: if methods for processing massive datasets were available, we could, for example, enable the online control of the activity of populations of neurons.

This talk is concerned with the data-processing methods that are necessary to make possible these closed-loop interactions in the specific context of retinal electrical stimulation (a proposed framework for the development of high-resolution retinal prosthesis). Here, the most elementary step is to single-out activity of individual neurons out of voltage recordings (spike sorting), in order to determine how cells respond to electrical stimuli. The main challenge is that recordings are heavily corrupted by stimulation-induced artifacts, which render the usual methods for spike sorting useless. We propose a method based on the modeling of the artifact as a Gaussian Processes with covariance given by a Kronecker product of non-stationary Kernels. Our method scales well to dense large-scale recordings and preliminary results show it leads to an accurate spike sorting algorithm.

Posted on Jan 12, 2016 in Seminario Aprendizaje de Máquinas, Seminars