Connectome-derived RNNs: task–region alignment

A brain region's wiring is matched to the computation that region evolved to perform — and that match carries over to artificial networks. We seed matched recurrent networks with connectomes from three Drosophila brain regions (optic lobe, mushroom body, central complex) and show that each region's wiring confers a task-specific advantage on its native task over size- and degree-matched random controls: structure→function alignment, demonstrated causally.

The effect is sharpest for the central complex on path integration — its connectome beats a uniform-random control and a degree-matched-random control (the degree-matched network is, if anything, worse than random), so the advantage is the region's specific wiring topology: not generic sparsity, not the degree distribution, not the synaptic weights. The optic lobe leads on optic flow and the mushroom body on associative recall, filling out the region × task alignment grid.

Cell = connectome's % advantage over its random control (sign-corrected, + = connectome better); black box = each region's native task. Per-cell numbers, controls, and statistics: docs/results/region_task_matrix/.

---

Repository map

path	what's here
src/	the library — connectome loading, control generation, models, training (src.train, src.connectome, …). All scripts import from here.
scripts/	entry points, grouped by topic — see scripts/README.md.
connectomes/	prepared connectome substrates (adjacency .npz + structure runs), used as --matrix inputs. (git-ignored data; regenerable via scripts/connectome/.)
data/	external datasets (DSEC flow, MNIST, Omniglot, larva). (git-ignored.)
outputs/	raw run artifacts (checkpoints, metrics), grouped under outputs/runs/<topic>/ — see outputs/README.md. (git-ignored.)
docs/	method writeups (docs/<topic>.md) and curated results with figures (docs/results/<experiment>/). Index: docs/results/README.md.
experiments/	experiment configs. flywire_cache/ raw connectome dumps. plumetracknets/ plume sub-project. tests/ unit tests.

scripts/ layout

connectome/ (build substrates) · flow/ (optic flow, DSEC) · mqar/ (associative recall) · associative/ (mushroom-body associative learning + benchmarks) · arbitrary/ (foreign-task battery) · path/ (central-complex path integration & dynamics) · continual/ · plume/ · classification/ · transfer/ · figures/ (plotting) · benchmarks/ · patent/ · setup/.

---

Quickstart

pip install -r requirements.txt          # GPU box setup: docs/aws_g7e_amazon_linux_setup.md

# regenerate the headline figure
python scripts/figures/plot_region_task_heatmap.py

# the central-complex / path result (strongest, degree-control-surviving cell)
python scripts/path/run_path_offdiagonal.py \
  --regions CX:connectomes/cx_polar_bump_seed0 --out-root outputs/runs/path/cx_deg \
  --seeds 0 1 2 --epochs 12 --train-count 8000      # models: connectome / random / weight_shuffle / degree_shuffle

# a task-specificity control (sequential MNIST — foreign to every region)
python scripts/arbitrary/run_arbitrary_tasks.py --task seq_mnist \
  --matrix connectomes/flywire_mushroom_body/adjacency_unsigned.npz \
  --models hemibrain_seeded weight_shuffle random_sparse no_recurrence --seeds 0 1 2

The original frozen-connectome benchmark CLI (run_benchmark.py) and its scientific notes are preserved in docs/cx_bpu_benchmark.md.

---

Key results — docs/results/<experiment>/

• region_task_matrix/ — the region × task alignment grid (figure above) + per-cell numbers.
• CX → path (degree-matched control) — the central complex's connectome beats random and degree_shuffle; degree-matched random is worse than random, so the advantage is the specific wiring pattern — not sparsity, degree distribution, or weights. The cleanest demonstration that region-specific structure encodes a task-aligned inductive bias.
• OL → optic flow — the optic-lobe connectome learns flow fastest, a region-aligned sample-efficiency gain on real DSEC event-camera data.
• MB → associative recall — the mushroom-body connectome reaches near-ceiling associative recall with a ~1.8× sample-efficiency boost over random.
• Task-specificity (controls) — the advantage is absent on tasks foreign to every region (sequential MNIST, arithmetic — with recurrence proven load-bearing), confirming it tracks the task, not network size.

Per-cell numbers, multi-seed statistics, and boundary cases: docs/results/region_task_matrix/README.md.

Controls vocabulary (used throughout)

connectome/hemibrain_seeded/connectome_bpu = real wiring · random/random_sparse = uniform random support · weight_shuffle = same topology, scrambled weights (isolates weights vs topology) · degree_shuffle/degree_preserving_random = random topology with matched degree sequence · no_recurrence = W_rec zeroed (proves recurrence is load-bearing).