Learning multiscale neural metrics via entropy minimization

摘要: In order to judiciously compare neural responses between repeated trials or stimuli, a well-suited distance metric is necessary. With multi-electrode recordings, response spatiotemporal pattern, but not all of the dimensions space and time should be treated equally. understand which input are more discriminative improve classification performance, we propose metric-learning approach that can used across scales. This extends previous work linear projection into lower dimensional space; here, multiscale metrics kernels learned as weighted combinations different on each response's dimensions. Preliminary results explored cortical recording rat during tactile stimulation experiment. Metrics both local field potential spiking data explored. The weights reveal important response, nearest-neighbor performance.