Temporal processing with plastic short term synaptic dynamics

Abstract

Event Abstract Back to Event Temporal processing with plastic short term synaptic dynamics Sensory processing such as speech and music perception, time interval estimation, or motion processing requires the central nervous system to extract and process information encoded in the temporal domain. Psychophysical studies of interval discrimination tasks have highlighted the brain's ability to decode temporal features of incoming stimuli over timescales ranging from tens-of-milliseconds to seconds. However, the neural mechanisms that underlie such processing on the basis of the spiking activity of sensory neural populations are unknown. Previous models have suggested that short-term synaptic plasticity is a key element in spike-based temporal computations[1]. These models have been limited by the lack of an efficient spike-based learning rule that is capable of tuning the dynamics of synaptic connections to the requirements of a given processing task. To overcome this limitation, we have derived a novel supervised spike-based learning rule for dynamical synapses[2], extending our recently proposed tempotron model[3]. By jointly acting on pre-synaptic release probabilities and post-synaptic response amplitudes as well as on synaptic recovery and facilitation time constants, this modification can change the dynamical properties of a given synaptic connection on the behaviorally important time scale of several hundred milliseconds. An `integrate-and-fire' model neuron with plastic dynamical synapses exhibits high capacity to decode spike-timing based information in the temporal domain. Moreover, by decoupling the processing of temporal information from the somatic voltage integration, plastic synaptic dynamics allow neurons to multiplex temporal features that arrive at different synaptic sites with high robustness to interference and neural noise. While the non-linear dependence of dynamic synaptic transmission on the input spike history allows the post-synaptic cell to process information in the temporal domain, the same history dependence limits a neuron's capability to process continuous streams of incoming spikes that prevent synapses from relaxing to their resting states. We demonstrate that this "reset" problem[1] can be overcome within populations of dynamical synapses. By appropriately aligning their dynamics, the effects of previous spikes can be effectively canceled. For instance a neuron can learn to detect a certain inter-spike interval independently of the preceding spikes. This active spike "forgetting" mechanism requires only one additional synapse per spike canceled. Importantly, this spike cancellation mechanism operates for arbitrary times of the previous spikes. As a result post-synaptic neurons can learn to process temporal information embedded within a continuous stream of inputs. A comparison between our model and the psychophysics of interval discrimination is carried out. Experimental studies have suggested a plethora of putative interaction sites between synaptic dynamics and long-term synaptic changes. Our new learning rule offers a computational rationale for such interactions. These results uncover powerful temporal computational capabilities of dynamical synapses within spiking neural networks. We suggest that training the synaptic dynamics could be mediated by long-term modifications in short-term synaptic plasticity in neocortex[4].