Adaptive probabilistic neural coding from deterministic spiking neurons: analysis from first principles (1111.0097v2)

Abstract: A neuron transforms its input into output spikes, and this transformation is the basic unit of computation in the nervous system. The spiking response of the neuron to a complex, time-varying input can be predicted from the detailed biophysical properties of the neuron, modeled as a deterministic nonlinear dynamical system. In the tradition of neural coding, however, a neuron or neural system is treated as a black box and statistical techniques are used to identify functional models of its encoding properties. The goal of this work is to connect the mechanistic, biophysical approach to neuronal function to a description in terms of a coding model. Building from preceding work at the single neuron level, we develop from first principles a mathematical theory mapping the relationships between two simple but powerful classes of models: deterministic integrate-and-fire dynamical models and linear-nonlinear coding models. To do so, we develop an approach for studying a nonlinear dynamical system by conditioning on an observed linear estimator. We derive asymptotic closed-form expressions for the linear filter and estimates for the nonlinear decision function of the linear/nonlinear model. We analytically derive the dependence of the linear filter on the input statistics and we show how deterministic nonlinear dynamics can be used to modulate the properties of a probabilistic code. We demonstrate that integrate-and-fire models without any additional currents can perform perfect contrast gain control, a sophisticated adaptive computation, and we identify the general dynamical principles responsible. We then design from first principles a nonlinear dynamical model that implements gain control. While we focus on the integrate-and-fire models for tractability, the framework we propose to relate LN and dynamical models generalizes naturally to more complex biophysical models.