by
Raoul-Martin Memmesheimer(Radboud University, Nijmegen)
→
Europe/Stockholm
FB51
FB51
Description
Precisely timed patterns of spikes have been experimentally found in several neural systems, ranging from electric fish to monkeys and humans. Since such patterns encode information, e.g. on touch signals and visual stimuli, they have been considered to be key elements of neural processing.
We investigate and develop supervised learning paradigms in networks of deterministic spiking neurons that allow to learn predefined precisely timed patterns of spikes in response to precisely timed input spike sequences. The task can be interpreted as a recoding of the input sequences within the framework of a timing-based neural code. In contrast to learning schemes which rely on the stochastic nature of neurons, the studied paradigms do not require a costly sampling of the probability distributions of voltage traces or spike sequences. We give proofs of convergence in finite time and estimate the network capacity numerically and analytically. Finally, we show that the paradigms can be generalized to learn spike patterns in recurrent neural networks.
The learning paradigms can be directly employed for the associative recall of temporal sequences as well as for the transfer of spike sequences between networks with different architectures, locations and properties, e.g. for the transfer from a short-term to a long-term memory storage location. Other applications include learning to predict events and to plan temporally precise actions. Examples are the generation of associations between a conditioned and an unconditioned stimulus in a classical conditioning task and the generation of timed conditioned responses. Finally, the supervised learning might be combined with stochastic synaptic dynamics to yield a reinforcement-type learning paradigm.
work done together with Ran Rubin, and Haim Sompolinsky
