Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Biologically Constrained Heterogeneity: Hybrid Spiking-Nonspiking Connectome-Constrained Models

Nalini Ramanathan¹, Maren Eberle¹, Erdem Varol¹; ¹New York University

Presenter: Nalini Ramanathan

In many neural circuits, information is processed through the coordinated activity of spiking and nonspiking neurons. While precise cellular parameters are often difficult to obtain, the classification of a neuron as graded or action-potential based is often well-characterized (e.g., photoreceptors tend to be nonspiking across organisms, whereas higher-order levels of the visual system such as RGCs in vertebrates and LPTCs in Drosophila tend to be spiking). We exploit this known heterogeneity to investigate whether directly connected spiking-nonspiking hybrid architectures provide a computational edge. We first validate the viability and robustness of this architecture on an auditory spiking identification task, the Spiking Heidelberg dataset, and find that hybrid networks outperform homogeneous baselines most strongly under limited neurons or training data. We then demonstrate the viability of this framework in a connectome-constrained model of the Drosophila antennal lobe which achieves 93-95% test accuracy when trained on a 200+ odor classification task and significantly worse accuracy (as low as 50%) for shuffled versions. When analyzed on principle components, SNN and Hybrid PN outputs are more linearly separable than RNN PN outputs. Given that PNs are observed to have linearly separable output for odor detection in biology and the loss of full information in transmission, this may suggest that spiking introduces greater biological realism from a linear separability perspective.

Topic Area: Methods, Tools, Theory & Neural Coding