Research map / Probabilistic and scientific modeling / Dynamical systems
Seemingly Redundant Modules Enhance Robust Odor Learning in Fruit Flies
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds a spiking fly olfactory circuit model (PN, LN, KC, MBON) and shows lateral inhibition helps odor discrimination at low/medium noise while spike-frequency adaptation dominates at high noise, and combining both gives the best accuracy.
Problem
Whether the two seemingly redundant sparsifying mechanisms in the fly olfactory circuit, lateral inhibition and spike-frequency adaptation, play redundant or distinct roles in learning odor discrimination under varying noise.
Approach
A LIF spiking model of the fly olfactory circuit with LI (LN-to-PN inhibition) and SFA (spike-triggered adaptation currents in PNs, LNs, KCs) trained with BPTT and surrogate gradients on KC-to-MBON weights, comparing the mechanisms across noise intensities.
Key results
- LI model reaches 98.30% accuracy at noise 0 for 1,000 classes vs 91.70% for baseline; at noise 0.3 SFA reaches 69.34% vs LI 53.82% and baseline 59.03%.
- SFA outperforms LI at N.I. >= 0.20 for 1,000 classes and >= 0.15 for 5,000 classes, consistent with noise-dependent recruitment.
- With OU noise at N.I. 1.5, SFA reaches 77.49% vs LI 69.72% and baseline 67.55% for 1,000 classes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add lateral inhibition (ablation) | Fly olfactory circuit · Gaussian noise, low- and medium-noise odor discrimination | accuracy: 6.80% when N.I. = 0.0; 18.52% when N.I. = 0.1 | Helped |
| Add lateral inhibition (ablation) | Fly olfactory circuit · Gaussian noise, high-noise odor discrimination | accuracy: -4.5% when N.I. = 0.30 | Hurt |
| Add spike-frequency adaptation (ablation) | Fly olfactory circuit · Gaussian noise, odor discrimination across tested noise levels | accuracy: 1.80% when N.I. = 0.0, 3.65% when N.I. = 0.1, and 11.56% when N.I. = 0.30 | Helped |
| Use LI and SFA together | Fly olfactory circuit · Gaussian noise, low- and medium-noise odor discrimination | accuracy | Helped |
| Use LI and SFA together | Fly olfactory circuit · Gaussian noise, high-noise odor discrimination | accuracy | Mixed |
| Add lateral inhibition (ablation) | Fly olfactory circuit · OU noise, 1,000-class odor discrimination | accuracy | Helped |
| Add spike-frequency adaptation (ablation) | Fly olfactory circuit · OU noise, 1,000-class odor discrimination | accuracy | Helped |
| Increase input signal strength | Fly olfactory circuit · 10,000-class odor discrimination, noise = 0.1 | accuracy: 21.80% to 96.87% | Helped |
Surprising findings
- LI at high noise (N.I. 0.30) impairs accuracy below the baseline, with the reduction smaller for stronger inhibition.
- Baseline (no sparsification) reaches 74.72% accuracy at 10,000 noise-free classes, suggesting the circuit's other features provide strong classification capacity.
Limitations
- Odors are artificially generated rather than natural odor stimuli.
- Training uses backpropagation through time, which is biologically implausible; ORN-KC weights are fixed.
- Parameter and noise ranges examined are limited.
Experimental setup
Datasets: synthetic odor dataset adapted from prior work (uniform prototypes with Gaussian or OU noise)
Benchmarks: synthetic odor discrimination (1000-10000 classes, Gaussian and OU noise)
Metrics: classification accuracy convergence epochs accuracy gain per epoch
Baselines: Baseline (no LI or SFA) LI model SFA model
Builds on: snnTorch FlyLoRA
Closest papers by meaning
- Inference with correlated priors using sisters cells (NeurIPS 2025 · similarity 0.82)
- Model Based Inference of Synaptic Plasticity Rules (NeurIPS 2024 · similarity 0.81)
- Spike-timing-dependent Hebbian learning as noisy gradient descent (NeurIPS 2025 · similarity 0.81)
- Shaping the distribution of neural responses with interneurons in a recurrent circuit model (NeurIPS 2024 · similarity 0.81)
- The Bayesian sampling in a canonical recurrent circuit with a diversity of inhibitory interneurons (NeurIPS 2024 · similarity 0.81)
- FEEL-SNN: Robust Spiking Neural Networks with Frequency Encoding and Evolutionary Leak Factor (NeurIPS 2024 · similarity 0.80)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.