Discriminating temporal patterns: spiking neurons and 'ideal observers' 

Haim Sompolinsky 

Animals need to detect or discriminate stimuli which are rich in temporal structure.  Our recently developed tempotron model demonstrates that an  integrate and fire neuron can learn to discriminate between stimuli which are encoded in spatio-temporal spike patterns. This raises the interesting question of how well simple neuronal circuits perform in temporal processing relative to an 'ideal observer'.

      A standard ideal observer model which is unlimited in its memory capacity may provide unrealistic bounds of performance for temporally restricted neural systems. I will describe a Dynamic Ideal Observer model which incorporates temporal locality and other dynamical-system features into the ideal observer decision process. This model provides useful bounds for the performance of simple neural systems in discriminating temporally extended stimuli.