On the Computational Power of Recurrent Circuits of Spiking Neurons

摘要

Understanding the structure of real-time neural computation in highly recurrent neural microcircuits that consist of complex heterogeneous components has remained a serious challenge for computational modeling. We propose here a new conceptual framework that strongly differs from all previous approaches based on computational models inspired by computer science or artificial neural networks. It is based on a rigorous mathematical model, the liquid state machine, whose computational power is analyzed in this article, both for the case of time -- varying analog input and for the case where the input consists of spike trains. The theoretical analysis implies that recurrent circuits are able to carry out complex real-time computations on such inputs, even several such computations in parallel, provided that they are able to separate different inputs through different activation patterns at subsequent time points. Furthermore, biologically realistic recurrent circuits of spiking neurons, consisting of heterogeneous neurons and synapses with different time constants, appear to be particularly good at this separation task. Based on this new approach one can now for the first time employ computer models for biologically realistic neural microcircuits as central processing units for complex computational tasks.