IPAB Workshop - 14/5/26

Title:

Connectome-Constrained Spiking Networks Reproduce Emergent Computations in Larval Olfactory Pathway

Abstract:

The larval Drosophila melanogaster connectome enables structurally faithful circuit models, but provides only wiring, not biophysical dynamics. We present a spiking neural network model of the larval olfactory pathway, from odor receptors via the antennal lobe to Kenyon cells in the mushroom body. Connectivity is fixed by the connectome and complemented by biologically plausible dynamics, with only 449 trainable parameters. The model is trained on a 28-odor classification task via two-phase ANN-to-SNN transfer learning with surrogate gradient descent under biophysical constraints. It spontaneously reproduces decorrelation in the mushroom body, APL suppression of KC activity by about half, concentration-invariant gain control across a 17-fold range, and sub-additive KC recruitment for odor mixtures. Ensemble training converges to consistent representations, demonstrating connectome dominance. Systematic ablation reveals individual circuit contributions and generates testable hypotheses. The connectome, combined with biophysical constraints and a simple learning objective, produces computations matching independent biological measurements, establishing a generalizable methodology for translating connectome data into testable spiking models.