MACHINE LEARNING APPROACHES FOR ELECTROPHYSIOLOGICAL CELL CLASSIFICATION

Carnegie-Mellon University, Pittsburgh PA

Investigators

Abstract

ABSTRACT We will use our expertise in somatosensory organization and plasticity to develop novel and automated solutions for cell identification based upon neural activity, in order to decode the algorithms neural circuits use for information processing. Extracellular recordings in sensory cortex have been thought to primarily represent excitatory neuron activity, since these cells comprise ~80% of the total cell population. However, targeted cell recordings in S1 reveal that firing activity is dominated by inhibitory neurons, and that excitatory neurons can show 10-100 fold lower firing rates depending on cortical layer. Furthermore, new findings that reveal the complex relationship between different subtypes of inhibitory neurons make it difficult to relate blindly-recorded firing activity to local- or network-level computations. Clearly, cell-types matter, and massively parallel extracellular recordings that do not enable the simultaneous identification of multiple cell types will be limited in identifying principles for information transmission and encoding. Based on our preliminary findings, we hypothesize that the complex spontaneous and evoked spike trains from molecularly-identified neurons will provide unique and cell-type specific signatures that will enable cell identification from in vivo extracellular recordings. In collaboration with computer scientists at Carnegie Mellon, we will develop machine-learning algorithms for cell classification, using data collected from in vitro and in vivo recordings.