Abstract
Across diverse organisms, the temporal dynamics of spiking responses between neurons, the neural synchrony, is crucial for encoding different stimuli. Neural synchrony is especially important in the insect antennal (olfactory) lobe (AL). Previous studies on synchronization, however, rely on pair-wise synchronization metrics including the cross-correlogram and cos-similarity between kernelized spikes train. These pair-wise analyses overlook an important aspect of synchronization which is the interaction at the population neuron level. There are also limited modeling techniques that incorporate the synchronization between neurons in modeling population spike trains. Inspired by recent advancements in machine learning, we leverage a modern attention mechanism to learn a generative normalizing flow that captures neuron population synchronization. Our method not only reveals the spiking mechanism of neurons in the AL region but also produces semi-interpretable attention weights that characterize neuron interactions over time. These automatically learned attention weights allow us to elucidate the known principles of neuron synchronization and further shed light on the functional roles of different cell types (the local interneurons (LNs), and projection neurons (PNs)) in the dynamic neural network in the AL. By varying the balance of excitation and inhibition in this neural circuit, our method further uncovers the pattern between the strength of synchronization and the ratio of an odorant in the mixture.
Author Summary The olfactory system can accurately compute the mixture of volatile compounds emitted from distant sources, enabling the foraging species to exhibit fast and effective decisions. However, altering ratios of one of the compounds in the mixture could be perceived as a different odor. Leveraging the current understanding of neural synchronization on sensory neural regions of insects, we construct a spatial-temporal attention normalizing flow, which partially replicates the AL region’s functionality by learning the spiking mechanics of neurons. Beyond providing insights of the spiking mechanism of neurons in the AL region, our method also produces semi-interpretable attention weights that characterize neuron interaction over time. These automatically learned attention weights allow us to dissect out the principles of neuron synchronization and interaction mechanisms between projection neurons (PNs) and local neurons (LNs). Utilizing our accurate model of these AL functionality, we show evidence that the behavioral relevant compounds are closely clustered together while varying the intensities of one of the behavioral compounds in the mixture could attenuate the synchronization
Competing Interest Statement
The authors have declared no competing interest.
Footnotes
↵* jriffell{at}uw.edu
