firefly — wildfire damage detection through a frozen fly brain

Can a Drosophila melanogaster connectome — never modified, never trained — act as the feature extractor of a binary image classifier? A satellite tile's 99-dim retinal sample (54 brightness + 45 red-dominance patches) drives the 99 sensory neurons of the MaleCNS v1.0 circuit (166,700 neurons, 25.6M frozen synapses, LIF dynamics); the 48-tick response is read as log1p spike counts of a fixed set of 1,024 downstream neurons; a tiny 64-unit MLP decoder — the only trained component — maps spikes to undamaged vs damaged.

Task framing: a plain ML benchmark on the Etkin satellite structure-damage tiles (binary collapse of 5 assessor damage grades). Numbers below are the committed benchmark; the protocol (split, downstream pick, decoder family, seeds, ablations) was committed before any binary result was generated.

Results

18,318 unique tiles (hash-deduped; 1,222 undamaged / 17,096 damaged — the majority class predicts 93.3% accuracy, so raw accuracy is meaningless here; balanced accuracy and ROC-AUC are the primary metrics). Same split, same frozen 1,024-downstream pick, same decoder family, 3 decoder seeds, 1000-resample bootstrap CIs in results/benchmark.json.

feature extractor	bal acc ↑	AUC ↑
chance	0.500	0.500
frozen fly circuit (the experiment)	0.588 [0.541–0.614]	0.622
shuffled connectome (degree-preserving rewiring)	0.585	0.625
all-excitatory (inhibition removed)	0.609	0.642
random sparse wiring (same nnz)	0.510	0.632
pooled pixels, no circuit	0.655	0.711
label shuffle (leakage check)	—	0.400

Verdict

1. The specific wiring carries no detectable task-relevant structure. The real connectome (AUC 0.622) is indistinguishable from a degree- preserving rewiring (0.625), from removing all inhibition (0.642), and from a random sparse matrix with matched nnz (0.632). Whatever lifts the score above chance lives in the retinal encoding + generic nonlinear mixing dynamics, not in the connectome's connectivity.
2. Pooled pixels beat the circuit (AUC 0.711 vs 0.622; balanced acc 0.655 vs 0.588) — the 48-tick fixed circuit is a lossy, noisy transform of its 99-dim drive.
3. Lesions have no specificity. Silencing the optic lobe, the central brain, the VNC, or a matched-size random neuron set all collapse the frozen decoder to the majority-class predictor (bal acc ≈ 0.50). The above-chance signal needs the whole circuit; it is not specifically visual. Removing the image entirely does the same.
4. No leakage detected (label-shuffle control AUC 0.400, i.e. the chance band; strong leakage would push AUC ≫ 0.5).

An interesting, decisive negative result: as driven here (99-patch retinal sample → 99 sensory neurons), the fly brain is interchangeable with a random fixed projection. What would falsify that conclusion: driving it with anatomically motivated inputs (e.g. columnar retina geometry, motion) or tasks closer to what the optic lobe computes.

Live demo (k3s NodePort, LAN): http://192.168.2.36:30181 — tile in, damage score + connectome wave out, lesion switches replay the ablation live, held-out hits/misses gallery included.

Method

tile --> retinal sample (99 patches) --> 99 sensory neurons (drive)
      --> frozen MaleCNS v1.0, 48 LIF ticks
      --> log1p spike counts, fixed 1,024 downstream neurons
      --> standardize --> 64-unit MLP (only trained part) --> P(damaged)

• Downstream pick: busiest non-sensory neurons by total spikes on the first 512 training drives (drive-derived only, never labels), frozen for every condition including controls.
• Split: stratified 85/15, seed 0, hash-deduped tiles. The source has no scene metadata, so near-duplicate scene leakage across the split is possible — a stated limitation, identical for every condition, so comparisons stand.
• Metrics: balanced accuracy + ROC-AUC at threshold 0.5 (scores not calibrated), 1000-resample bootstrap CIs, 3 decoder seeds.
• Controls: shuffled (degree-preserving rewired) connectome, all-excitatory (inhibition removed), random sparse wiring, pooled-pixel MLP on the same retinal sample, label-shuffle leakage check.

Run

python3 -m firefly.train --device cuda      # trains, benchmarks, writes results/
python3 -m uvicorn serve.app:app --port 8000  # FIREFLY_OUT=/library/datasets/firefly
kubectl apply -f deploy/namespace.yaml -f deploy/web.yaml   # NodePort 30181

Training reads /library/datasets/fire_sat/etkin/*.parquet and the processed MaleCNS cache (/library/datasets/malecns_v1/processed-firefly/); serving reads only readout.npz, results/benchmark.json, samples/ from FIREFLY_OUT — ~1 s per tile on CPU.

What this is not

Not biologically validated vision: the retinal mapping is an engineering choice, dynamics are doomfly-style approximations, and pooled pixels beat the circuit (see table) — the question here is whether connectome wiring carries task-relevant structure, not whether a fly brain is a good camera.

Credits

MaleCNS v1.0: HHMI Janelia + Google (CC-BY) · data: Etkin et al. via kevincluo/structure_wildfire_damage_classification (apache-2.0) · recipe after jerryjliu/fly_ocr (MIT) and nftechie/doomfly (MIT). Companion playground with the game-based demos: seanphan/flyt3.