Speaker: Nikolai Chapochnikov, Flatiron Institute

Title: Neural computations for stimulus whitening by the olfactory system

Abstract: Most sensory systems encode external stimuli using multidimensional patterns of neural activity. To disentangle distinct stimuli that elicit similar patterns of activity, sensory systems often perform pattern decorrelation to facilitate downstream processing. However, how exactly decorrelation is implemented by neural circuits is not entirely understood. Here, we study the peripheral olfactory neural circuit of the drosophila larva, which is thought to implement decorrelation. The circuit is composed of 21 olfactory receptor neurons (ORNs) connected through a feedback loop with an ensemble of inhibitory local neurons (LNs). We hypothesized that stimulus decorrelation would rely on synaptic connectivity that is tuned to the statistics of ORN activity. Indeed, using data from connectomics and ORN responses to an ensemble of odors, we show that the ORNs -> LNs synaptic weight vectors are significantly aligned with the patterns of ORNs activity. In addition, we show the subspace of top principal vectors of ORN activity significantly overlaps with the subspace of ORNs -> LNs connection weight vectors. To understand the computation performed by such a circuit architecture and the mechanisms that could lead to the experimentally observed synaptic weights, we used a mathematical description of the circuit dynamics and Hebbian synaptic learning. Our analysis demonstrates that such a circuit whitens the ORN activity patterns by dampening top principal components via the negative feedback from LNs. We also show that through unsupervised learning the synaptic weights self-organize so that the subspace created by the connectivity vectors aligns with the ORN principal activity subspace, as observed in our data analysis. In summary, we can explain the observed relationships between synaptic weight vectors and neuron activity and relate it to the computational task of whitening. Our study thus supports the idea that the larva olfactory neural circuit learns and dampens through negative feedback the most statistically prevalent direction of ORN activity to enable better downstream odor differentiation.