Authors : Veronika Koren and Stefano Panzeri
High-level description Neurons communicate with each other by sending to each other brief electrical impulses that we call spikes. In the last decades, outstanding research on spiking networks has brought crucial insights into the dynamics of these complex interacting systems. We however still lack the link between neural dynamics and neural function, where the neural function is presumably to generate neural signals that underlie perception, thinking, decision-making and motor behaviour. This link is essential, as generation of behaviorally relevant signals is the purpose of neural networks in biological brains.
Our work extends previous theoretical work on efficient coding with spikes (Boerlin et al. 2013; Koren and Deneve, 2017) to the biologically relevant neural architecture with excitatory and inhibitory neurons. We suggest an efficient population code where every spike is fired to improve the internal representation of the stimulus of neuron's population, and the activity is constrained by limited metabolic resources (i.e., a cost on high firing rates). We obtain a generalised leaky integrate-and-fire neuron model for both E and I cell types, a model of neural dynamic that is known to accurately describe the spiking activity of biological networks. Besides obeying Dale’s law, our efficient E-I spiking network gives the account of neuron-intrinsic currents such as spike-triggered adaptation. Our network allows linear transformation of the input features that is executed through specific connectivity structure between excitatory (E) neurons.
We find that the complexity of input transformation determines the rank of E-E conenctivity and generates qualitatively different neural dynamics. With rank higher than 1 and in presence of sufficient noise in the membrane potential of single neurons, the network generates slow rhythmic spontaneous activity similarly to what is observed in cortical networks during sleep.
Published in: Advances in Neural Information Processing Systems 35 (NeurIPS 2022) | to NeurIPS