Biomimetic Sparsity vs. Dense Computation: Benchmarking the Whole-Brain Drosophila Connectome Against Scale-Matched Artificial Neural Networks Under Strict Resource Constraints
Year: 2026
Type: preprint
Abstract: Recent breakthroughs in whole-brain connectomics—culminating in the adult fruit fly (Drosophila melanogaster) full-brain wiring diagram resolved by the FlyWire consortium (139,255 neurons and ~54.5 million synapses)—provide the first synaptic-resolution computational blueprint of an autonomous organism. In sharp contrast to biological nervous systems that operate under strict metabolic ceilings, contemporary deep artificial neural networks (ANNs) rely on dense, synchronous matrix multiplications that incur substantial energy and memory footprints. In this work, we introduce a rigorous, reproducible benchmark comparing a connectome-constrained bio-mimetic Spiking Neural Network (Bio-SNN) against an identically sized Dense Artificial Neural Network (Dense-ANN) under strict resource constraints and identical behavioral demands in a turbulent odor plume environment. Our empirical results demonstrate: (1) The Bio-SNN achieves comparable goal-directed trajectory fidelity with a 1,997.7x reduction in computational operations (5.65 x 10^5 SynOps vs. 1.13 x 10^9 FLOPs), translating to a 10,210.4x decrease in estimated thermodynamic expenditure (0.509 uJ vs. 5,194.23 uJ); (2) Under progressive random node ablation (0% to 50% lesion), the Bio-SNN exhibits graceful degradation due to its heavy-tailed log-normal synaptic topology, whereas the Dense-ANN exhibits severe performance collapse; and (3) Inherent Noise Attenuation: Event-driven spike integration filters high-frequency sensor noise without explicit parameter re-training.
Author: Vatansever, Rüzgar
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236696164
