Passive Nonlinear Dendritic Interactions as a Computational Resource in Spiking Neural Networks

摘要

Nonlinear interactions in the dendritic tree play a key role in neural computation.Nevertheless, modeling frameworks aimed at the constructionof large-scale, functional spiking neural networks, such as the NeuralEngineering Framework, tend to assume a linear superposition of postsynapticcurrents. In this letter, we present a series of extensions to theNeural Engineering Framework that facilitate the construction of networksincorporating Dale’s principle and nonlinear conductance-basedsynapses. We apply these extensions to a two-compartment LIF neuronthat can be seen as a simple model of passive dendritic computation.We show that it is possible to incorporate neuron models with inputdependentnonlinearities into the Neural Engineering Framework withoutcompromising high-level function and that nonlinear postsynapticcurrents can be systematically exploited to compute a wide variety ofmultivariate, band-limited functions, including the Euclidean norm, controlledshunting, and nonnegative multiplication. By avoiding an additionalsource of spike noise, the function approximation accuracy of asingle layer of two-compartment LIF neurons is on a par with or evensurpasses that of two-layer spiking neural networks up to a certain targetfunction bandwidth.