Seemingly Redundant Modules Enhance Robust Odor Learning in Fruit Flies

TL;DR

The study investigates why seemingly redundant sparsification mechanisms exist in the fruit fly olfactory circuit by modeling LI and SFA within a cerebellum-like pathway. A spiking network with large PN→KC expansion and sparse KC→MBON readout demonstrates that LI enhances odor discrimination in low to medium noise, while SFA provides robust gains across all noise levels, with SFA dominating under high noise. When combined, LI and SFA yield additive improvements, supporting the idea that biologically redundant modules are recruited contextually to optimize learning in complex environments. The work bridges neuroscience and neuromorphic design, illustrating how noise-adaptive recruitment of distinct sparsification strategies can improve robust learning in sensory circuits.

Abstract

Biological circuits have evolved to incorporate multiple modules that perform similar functions. In the fly olfactory circuit, both lateral inhibition (LI) and neuronal spike frequency adaptation (SFA) are thought to enhance pattern separation for odor learning. However, it remains unclear whether these mechanisms play redundant or distinct roles in this process. In this study, we present a computational model of the fly olfactory circuit to investigate odor discrimination under varying noise conditions that simulate complex environments. Our results show that LI primarily enhances odor discrimination in low- and medium-noise scenarios, but this benefit diminishes and may reverse under higher-noise conditions. In contrast, SFA consistently improves discrimination across all noise levels. LI is preferentially engaged in low- and medium-noise environments, whereas SFA dominates in high-noise settings. When combined, these two sparsification mechanisms enable optimal discrimination performance. This work demonstrates that seemingly redundant modules in biological circuits can, in fact, be essential for achieving optimal learning in complex contexts.

Figures (6)

• Figure 1: Schematic of the fly olfactory circuit model. Odor inputs are sensed and encoded by ORNs. After passing through the ORNs, odor-triggered spikes in PNs can be shaped by two factors: LI caused by LNs and SFA by inherent adaptation currents. Neuronal spikes in KCs may subsequently show different variation patterns to favor odor discrimination in MBONs.
• Figure 2: Odor discrimination accuracy for fly olfactory circuit variants in a noise‑free setting. Performance is shown for three configurations—--Baseline, LI, and SFA models—--as a function of the number of odor classes (1,000–10,000).
• Figure 3: Changes in discrimination performance of the SFA and LI models relative to the Baseline model across varying degrees of inhibition and adaptation under different noise intensities. Results are shown for 1,000-class odor discrimination only. The "Low," "Medium," and "High" conditions (distinguished by color) represent increasing strengths of the respective mechanisms, achieved by systematically adjusting the relevant synaptic weights: $w_{\text{LN} \to \text{PN}}$ in Eq. \ref{['eq:LI']} for LI, and $w_{\text{SFA,X}}$ in Eq. \ref{['eq:SFA']} for SFA. For both mechanisms, "Low," "Medium," and "High" correspond to weight increases in an approximate 1:2:3 ratio.
• Figure 4: Discrimination performance of the SFA model, LI model, Full (SFA + LI) model, and Baseline model under different noise intensities. The models were tested on odor discrimination tasks with 1,000, 2,000, and 5,000 odor classes.
• Figure 5: Impacts of LI and SFA on model convergence speed. The top (bottom) panel shows the number of training epochs required for convergence (the accuracy gain per epoch) for the Baseline, LI, and SFA models, plotted against noise intensity for different odor category sizes: 1,000, 2,000, and 5,000 classes. To avoid potential misinterpretations from relying solely on maximum accuracy and to objectively assess learning progress, convergence is defined as the point at which model accuracy growth plateaus. Specifically, a model is considered converged when the average improvement in accuracy over $n = 10$ consecutive epochs falls below a predefined threshold (threshold = $0.003$).