A benchmark for comparing connectome-topology-constrained neural architectures
derived from a real Drosophila nervous system against conventional AI models
on controlled computational tasks.
Most comparisons between "bio-inspired" and conventional AI architectures use approximate or synthetic connectivity data. NeuroWeave asks what happens when you use the actual wiring of a real nervous system as a structural constraint on a neural network — and test it rigorously against conventional AI under identical conditions.
It is an engineering experiment, not a neuroscience simulation. The goal is to find out whether real biological connectivity contains computationally useful structure — and to be honest about when it does not.
Does using the real wiring diagram of a fruit fly nervous system as a neural network architectural constraint produce measurable computational advantages over conventional AI baselines on controlled tasks?
The answer so far: it depends on the task.
MaleCNS v1.0 is the first complete connectome of the adult male Drosophila melanogaster central nervous system, produced by the FlyEM Project Team at HHMI Janelia Research Campus.
- ~21,000 neurons mapped
- Tens of millions of synaptic connections
- Neurotransmitter predictions per neuron
- Publicly released under CC-BY 4.0
A connectome is a structural wiring diagram — it records which neuron connects to which, how many synapses join them, and in which direction signals travel. It is not a model of what the neurons compute; it is a measurement of how they are connected.
NeuroWeave uses this connectivity data as an architectural mask on a neural network layer: only biologically real connections are allowed to carry learned weights. All other weights are structurally zeroed.
MaleCNS v1.0 connectome
│
▼
Extract subgraph
(150 neurons, 3,029 edges)
│
▼
Build connectivity mask
(sparse adjacency from real data)
│
▼
Constrained neural network layer
(weights only where biology says connections exist)
│
▼
Train on benchmark tasks
│
▼
Compare against:
├── Random graph (same density)
├── Dense MLP
├── LSTM
└── Random baseline
│
▼
Evaluate on held-out test set
(5 seeds, bootstrap 95% CI)
| Architecture | Accuracy |
|---|---|
| MaleCNS-derived | ~99% |
| Random graph (ER) | ~99% |
| MLP | ~99% |
| LSTM | ~99% |
| Random baseline | ~50% |
All trained architectures approach ceiling accuracy on static binary discrimination. T-001 does not discriminate between architectural families — an expected and scientifically valid negative result.
Confirmed results — 5-seed evaluation on held-out test set:
| Architecture | Mean accuracy | 95% CI | Trainable params |
|---|---|---|---|
| MaleCNS-derived + SA-010 | 94.8% | ±6.2% | 6,514 |
| MaleCNS-derived (frozen) + SA-010 | 72.2% | ±11.4% | 0 |
| LSTM | 57.6% | ±8.9% | 266,626 |
| Random baseline | 50.0% | — | 0 |
On the temporal memory task, the MaleCNS-derived model achieves 94.8% accuracy using 41× fewer trainable parameters than the LSTM (57.6%).
Scientific caveats — read before citing:
- The SA-010 configuration (decay 0.05 / leak 0.02) was identified through exploratory sensitivity analysis, then confirmed on a separate held-out test set.
- Confidence intervals are wide (±6–11%). The result is meaningful but not yet decisive at large scale.
- This result is task-specific. It does not generalise to other tasks, other subgraphs, or other organisms.
- This is not evidence that the fly brain is computationally superior.
The Live Arena is an interactive benchmark replay.
What it is: A browser demonstration that runs trial-by-trial replays of the T-002 delayed-recall task. Each trial draws its outcome from the measured benchmark accuracy probabilities:
- MaleCNS-derived + SA-010: 94.8% correct per trial
- LSTM: 57.6% correct per trial
What it is not: The trained PyTorch models are not running in the browser. No model weights are loaded client-side. The demo uses a seeded deterministic PRNG to generate trial outcomes consistent with the real benchmark statistics.
The caveat is permanently displayed on the arena page.
e:\animal-brain-benchmark\
├── abb/ # Python research engine
│ ├── data/ # MaleCNS data loading, graph construction
│ ├── models/ # Architecture implementations (GNN, MLP, LSTM, SA-010)
│ ├── tasks/ # T-001, T-002 task definitions
│ ├── sim/ # Simulation / evaluation loop
│ └── config/ # YAML experiment configs
├── scripts/
│ ├── tasks/ # Experiment runner scripts
│ ├── data/ # Data pipeline utilities
│ └── extract_graph_stats.py
├── artifacts/ # Real experiment outputs (JSON)
│ ├── campaign_1/ # T-001 multi-architecture results (10 architectures)
│ ├── campaign_2/ # T-002 stateful architecture results
│ ├── campaign_3/ # T-002 SA-010 5-seed results
│ ├── confirmation_t002/ # T-002 confirmed results (5-seed, held-out)
│ └── sensitivity_t002/ # T-002 sensitivity sweep across SA-010 params
├── configs/ # YAML experiment configurations
├── data/ # Raw MaleCNS parquet data
├── docs/ # Methodology, scope, assumptions documents
└── frontend/ # React/TanStack web interface
| ID | Description | Trainable params | Notes |
|---|---|---|---|
| A1-BIO | Connectome-constrained, trained | 6,514 | Primary model |
| A1-FROZEN | Connectome-constrained, frozen weights | 0 | Topology-only ablation |
Both models use the same 150-neuron subgraph with 3,029 directed edges. SA-010 leaky/decay temporal dynamics are added for T-002.
SA-010 assumptions (engineering, not biological fact):
- Decay rate: 0.05 / Leak rate: 0.02
- Selected via exploratory sensitivity analysis on T-002
- Not derived from membrane potential measurements
| ID | Description |
|---|---|
| A3-ER | Erdős–Rényi random graph (same edge density) |
| ID | Description | Trainable params |
|---|---|---|
| A0-RANDOM | Untrained random weights | 0 |
| A7-MLP | Multi-layer perceptron | varies |
| A8-LSTM | Long short-term memory | 266,626 |
Type: Static supervised classification · D_obs: 16 · Memory required: No
A single binary pattern is presented. The model classifies it in one forward pass. Baseline sanity check — establishes a performance ceiling.
Type: Working memory / temporal classification · D_obs: 16 · Memory required: Yes
A binary stimulus is shown at time step 0. Several gap steps follow with no relevant signal. At the end of the gap, the model must recall which stimulus it originally saw. A feedforward network cannot solve this without memory.
| Limitation | Detail |
|---|---|
| Prototype scale | 150-neuron subgraph is a small sample of ~21,000 neurons |
| Single organism | Results apply to one male Drosophila; not generalisable |
| Task scope | Only T-001 and T-002 completed; T-003/T-004 not yet run |
| SA-010 assumptions | Decay/leak parameters are engineering choices, not measured biology |
| Wide CIs | 5 seeds; 95% CI ≈ ±6–11% — results need larger-scale confirmation |
| Single subgraph | Only one 150-neuron subgraph tested |
This project does not claim:
- The model simulates a living fly brain
- Biological intelligence has been reproduced
- Results generalise to vertebrate or human brains
- The MaleCNS topology is universally superior to conventional architectures
python >= 3.10
pytorch >= 2.0
torch-geometric
neuprint-python # for data pipeline only; raw parquet included in data/git clone https://github.com/Titanium-xd/neuroweave.git
cd neuroweave
pip install -e .# PowerShell
$env:PYTHONPATH='.'
python scripts/tasks/run_confirmation_t002.py
# bash
export PYTHONPATH=.
python scripts/tasks/run_confirmation_t002.pyResults are written to artifacts/confirmation_t002/.
cd frontend
npm install
npm run dev
# → http://localhost:8080| Layer | Technology |
|---|---|
| Graph data | MaleCNS v1.0 parquet via neuprint-python |
| GNN | PyTorch Geometric — custom sparse linear operator |
| Temporal dynamics | Custom SA-010 leaky/decay layer (PyTorch) |
| Evaluation | 5-seed, held-out test split, bootstrap 95% CI |
| Frontend | React 19 + TanStack Router + TanStack Start |
| Data layer | JSON artifact files read by src/data/benchmark.ts |
| Deployment | Cloudflare Pages (Nitro, cloudflare-module preset) |
This project uses MaleCNS v1.0 produced by the FlyEM Project Team at HHMI Janelia Research Campus, released under CC-BY 4.0.
FlyEM Project Team et al.
"Sexual dimorphism in the complete connectome of the Drosophila
male central nervous system"
Cell, September 2026
https://male-cns.janelia.org
https://neuprint.janelia.org (dataset: male-cns:v1.0)
License: CC-BY 4.0 — https://creativecommons.org/licenses/by/4.0/
Any use of NeuroWeave that reproduces or derives from MaleCNS data must include the above attribution per the CC-BY 4.0 license terms.
- Scale: Test larger subgraphs (500, 2000 neurons) — does the T-002 advantage persist?
- Task diversity: T-003 (navigation), T-004 (lesion robustness), continuous control tasks
- SA-010 ablation: Isolate the contribution of connectome topology vs. leaky/decay dynamics
- Baseline breadth: Add transformers, state-space models (Mamba), liquid neural networks
- Multiple subgraphs: Test ≥3 subgraphs per experiment to assess subgraph sensitivity
- Reproducibility package: Docker image, locked dependency manifest, full seed logs
- NeuroWeave code: MIT License
- MaleCNS-derived data products: CC-BY 4.0 (attribution required)
Case Study by Parva Trivedi
NeuroWeave — Animal Brain Benchmark v1.0