Project 04Read · experimentPyTorch · CPU float32 · preregistered locally

Does fly wiring help a network predict text?

Take 512 real neurons and their 3,061 directed connections from the connectome. Make each connection a trainable weight, feed bytes of Shakespeare through learned embeddings, and train a readout. Compare against the same graph with its edges rewired while preserving every neuron's in- and out-degree. If the anatomy carries useful structure, the fly topology should win. It did not.

Graph

512 neurons · 3,061 edges · sampling seed 101

Confirmation matrix

35 runs · 5 seeds · 20,000 updates each · 4.46 h CPU

Primary effect

−0.0103 BPB, 95% CI [−0.0206, +0.0017]

Verdict

No fly-wiring advantage established
Figure: final test performance per condition and the primary paired comparison, whose interval crosses zero
Figure 01. Held-out test bits per byte for all 35 runs (hollow circles) and their means (diamonds); the predeclared paired effect with its 10,000-resample bootstrap interval. The dashed line is the +0.02 BPB practical threshold chosen before training.

The result

On the sealed test split, averaging the three rewired controls for each seed and subtracting the fly model: mean −0.0103 bits per byte, 95% paired bootstrap interval [−0.0206, +0.0017]. Four of five seeds favoured rewiring. The interval includes zero, so no clear disadvantage is established either, and it falls well short of the predeclared +0.02 BPB practical advantage. This is one subgraph from one specimen, mostly visual neurons, with about 92% of those neurons' filtered connectivity falling outside the subset. It is a null result for this experiment, not a result about biological learning.

ConditionRunsActive paramsTest BPBByte perplexityAccuracy
Fly topology552,9812.63886.2346.92%
Degree-rewired control1552,9812.62866.1847.11%
Random-density control552,9812.58275.9947.65%
Parameter-matched GRU553,0002.55755.8949.41%
No cross-node recurrence550,4322.67386.3845.92%

Validation-selected checkpoints scored once on 55,769 test positions. Train-only references: unigram 4.8506 BPB, bigram 3.6084 BPB. Lower is better.

Figure: all five matched seed comparisons and all 15 rewired test scores
Figure 02. Each seed's fly run against its three independently rewired controls.

The design

Figure: the actual 512-node fly graph and a rewired control, drawn in a common layout, with their adjacency matrices
Figure 07. The real 512-node subgraph and one rewired control in a common abstract layout, with directed adjacency matrices. Layout is not anatomy.

What was checked

After every stage an audit script replayed the saved validation scores and matched them to the logs within 10⁻⁶ BPB, verified code, plan, corpus and graph hashes, confirmed exact labelled degree preservation in the controls, matched initial edge-slot weights between fly and null arms, checked that every run consumed its full budget, and confirmed that no test loader had been opened. The final audit re-derived the paired bootstrap independently. The unit suite is 31 passed, 2 skipped (no CUDA or MPS in the deterministic CPU configuration).

Reproduce

cd projects/connectome-lm
python3.12 -m venv .venv && .venv/bin/pip install -r requirements.lock && .venv/bin/pip install --no-deps -e .
./fly-lm download --cache-dir ../../data/flat-connectome      # MaleCNS flat files, ~1.07 GB
./fly-lm prepare --nodes 512 --sampling-seeds 101
./fly-lm smoke                                               # 5 runs, minutes
./fly-lm run --config pilot                                  # 15 runs
./fly-lm tune --config pilot --output configs/development-selection.json
./fly-lm freeze --config confirmation --selection configs/development-selection.json --output configs/confirmation-frozen.json
./fly-lm run --config confirmation --freeze configs/confirmation-frozen.json   # 35 runs, ~4.5 h on a laptop CPU
./fly-lm final-evaluate --freeze configs/confirmation-frozen.json

The confirmation reports, final review, audits, per-run values and the figure source are in reports/confirmation-211428f1ca1f/. Checkpoints and prepared graphs are not in the repository.