Efficient computing of high-dimensional neural representations with biologically plausible E-I spiking networks
Authors: Veronika Koren
Presentation type: Short talk at SNUFA 2023 online workshop (7-8 Nov 2023)
A primary goal of neuroscience is to understand how neural computations are realized within the constraints and properties of biological networks. The framework of efficient coding with spikes previously suggested a theory of how neural networks might compute behaviorally relevant signals, but did not account for many biophysical properties of biological networks. In this study, we analyse the behavior of functional spiking networks of excitatory and inhibitory neurons, analytically derived from principles of efficient coding. We assumed that excitatory and inhibitory neurons follow separate and different coding objectives. We found that efficient coding is realized by a particular class of E-I spiking models of generalized leaky integrate-and-fire neurons with structured recurrent connectivity, where the connection strength is a rectified linear (ReLu) function of similarity in stimulus selectivity. By allowing different time constants at the population and single neuron levels, our model provides a computational account of local currents such as spike-frequency adaptation. The global state of the network is modulated by the metabolic cost on spiking that regulates the mean of the external current to the network. The structure in recurrent connectivity has important consequences on network behavior, as it results in stimulus selectivity in I neurons, precise average and time-dependent E-I balance, and lateral inhibition between excitatory units with similar selectivity. With balanced feedforward inputs, the optimal E-I spiking network is an unbiased estimator of a set of simultaneously evolving stimulus features and its accurate performance generalizes well over a wide range of time scales of the stimulus. The network is robust to added random heterogeneity in synaptic connectivity and displays plausible spiking activity for any reasonable set of parameters. The optimal set of parameters coincides with empirically measured parameters in primary cortical areas.
Watch it on Youtube