Deconstructing Odorant Identity via Primacy in Dual Networks

TL;DR

This paper introduces a neural network model inspired by the olfactory system that encodes odorant identity based on the strongest receptor responses, ensuring invariance to stimulus intensity and nonlinearities.

Contribution

It proposes a novel dual network approach that uses relative receptor responses to reliably identify odors, with a neural implementation based on Lagrange multipliers.

Findings

The strongest receptor responses suffice for odor identification.

A neural network can perform elastic net minimization for sparse stimulus recovery.

Odor identity can be decoded via dual computations in the piriform cortex.

Abstract

In the olfactory system, odor percepts retain their identity despite substantial variations in concentration, timing, and background. We propose a novel strategy for encoding intensity-invariant stimuli identity that is based on representing relative rather than absolute values of the stimulus features. Because, in this scheme, stimulus identity depends on relative amplitudes of stimulus features, identity becomes invariant with respect to variations in intensity and monotonous non-linearities of neuronal responses. In the olfactory system, stimulus identity can be represented by the identities of the p strongest responding odorant receptor types out of a species dependent complement. We show that this information is sufficient to recover sparse stimuli (odorants) via elastic net loss minimization. Such a minimization has to be performed under constraints imposed by the relationships between stimulus features. We map this problem onto the dual problem of minimizing a functional of Lagrange multipliers. The dual problem, in turn, can be solved by a neural network whose Lyapunov function represents the dual Lagrangian. We thus propose that networks in the piriform cortex compute odorant identity and implement dual computations with the sparse activities of individual neurons representing the Lagrange multipliers.