A manuscript has been submitted for this project and a pre-print will be coming soon!!
Figure: (a) Schematic of a feedforward spiking neural network capturing signal convergence and divergence in the moth visuomotor pathway (b) spike trains at each layer (c) input decodability (top) and mutual information (bottom) computed from either spike counts (magenta) or spike timing (yellow).
Previous experimental studies suggest that neural systems utilize mixtures of rate coding (where information is encoded in the average number of spikes within a time window) and temporal coding (where information is encoded in the precise timing of spikes). For example, cortical neurons are known to utilize rate coding, while at sensory peripheries and motor outputs, temporal coding is employed. What determines the optimal coding strategies in spiking networks of neurons, however, is unknown. In collaboration with the Sponberg lab, we aim to address this question and unveil how optimal coding is shaped by the underlying network structure. We hypothesize that rate coding is utilized when signals are projected onto a higher-dimensional space (signal expansion) while temporal coding has more advantage when signals are projected onto a lower-dimensional space (signal compression). To test this hypothesis, we built a feedforward network model of spiking neurons replicating the signal pathway of the moth visuomotor system including the initial divergence of sensory signals projected onto the brain, the information bottleneck at the VNC, followed by the secondary expansion and the final convergence to motor outputs. By computing mutual information and performing decoding analyses for time-varying input stimuli based on either spike counts or spike timing, we investigate how the sequence of convergence and divergence in the network determines coding schemes based on neuronal spikes. Our preliminary results show that each layer along the feedforward spiking network has alternating optimal coding strategies, and the temporal code indeed has a better representation of sensory information at layers where signals converge. In the next funding year, we will study how the two competing coding strategies encode motor output in addition to sensory input at different layers of the spiking neural network model and test our model predictions on experimental data collected from different stages along the moth sensorimotor pathway. Furthermore, we will systematically perturb the number, sequences, and ratios of expansion/compression motifs in the network to identify key structural features that shape optimal coding.