Recurrent Neural Network Model of the Olfactory System Develops Primacy and Predictive Coding

Olfaction is a sensory modality that plays an essential role for the survival of most animals. Yet, understanding the neuronal architecture of the olfactory system is incomplete. As a result, there is no complete description of the algorithms responsible for the computation of odor identity (what type of molecules is present in the environment) and intensity (how many molecules are present). Here we model the olfactory system as a recurrent neural network and attempt to derive these algorithms from the first principles. To this end, we present the network with simulated odorant mixtures and modeling the changes of concentration in the nose during natural sniffing. The goal of the network is to compute, in real time, the known odor identity and intensity. We trained the network to do so using backpropagation to derive an optimized algorithm which implements this task. Our model spontaneously develops an early recognition of both identity and intensity, also known as primacy coding. This implies that the network recognizes input stimuli early in the sniff cycle. This feature is not enforced on the network and emerges as a solution to the problem of identifying smell by the network ab initio. The model also shows activity patterns reminiscent of predictive coding. Overall, we used a deep learning model as an inference tool to investigate candidate algorithms of the olfactory system and identified two mechanisms which may be involved in recognizing olfactory stimuli.