Title: “State-dependent processsing with Spiking Neural Networks”
Abstract: Many different aspects of cognitive function express themselves as structured temporal sequences. On the other hand, several important organizational principles of the neocortex appear to imply a strong predisposition to acquire this temporal structure in a completely incidental/unsupervised manner.
In this work, we have explored the processes involved in implicit, structured sequence learning in biologically-inspired architectures in order to evaluate the character of on-line processing memory and finite precision computation in systems where the current state continuously interacts with and modifies the processing characteristics. We have demonstrated a prominent role of synaptic plasticity (particularly of inhibitory synapses) in representational and rule-guided learning, an effect achieved by maintaining compact dynamic representations and sparse, distributed activity patterns. We have highlighted a form of sequential metastability as a potential mechanism for sequence learning in neocortical circuits. In addition, we discuss how innate constraints in the patterning of the synaptic machinery throughout the neocortex may bias a circuit’s intrinsic timescales and memory capacity, while the high degree of complexity and heterogeneity may serve important computational purposes by expanding the circuit’s functional space.