Scienceexplained simply.
A visual story. No neuroscience background required. Nine chapters from a real nervous system to a controlled AI experiment.
Start with the wiring.
Scientists painstakingly mapped every neuron and every connection in the nervous system of an adult male fruit fly (Drosophila melanogaster). MaleCNS v1.0 contains roughly 21,000 neurons and tens of millions of synaptic connections — a real wiring diagram for a real nervous system.
Node positions are schematic — not anatomical coordinates. Real anatomical data is not used in the current prototype.
We can't run the whole thing in one experiment.
21,000 neurons is computationally demanding for an initial benchmark. We extract a smaller prototype subgraph — a connected subset of 150 neurons that preserves the wiring structure of the full connectome in a manageable form.
Graph statistics are real — extracted directly from the MaleCNS parquet data.
Now give the wiring something to do.
We use the connectome's wiring as a computational architecture — a layer in a neural network where only biologically real connections are allowed to carry signals. Inputs propagate through the connectome's topology and produce an output.
This is a computational model, not a simulation of biology. No neurons fire. No membrane potentials are computed. It is a mathematical graph with learned parameters, constrained to match real connectivity.
Think of it as: "what if a neural network's wiring was forced to look like this specific nervous system?"
Same problem. Same rules.
To test whether biological wiring provides any computational benefit, we compare it against conventional AI architectures under identical conditions. Same data. Same task. Same evaluation. Multiple random seeds to prevent cherry-picking.
- Identical training data per seed
- Identical held-out test split
- Same input dimension (D_obs = 16)
- 5 seeds per architecture
- Same evaluation metric
- Connectivity structure
- Trainable parameter count
- Temporal dynamics assumptions
The easy test wasn't enough.
The first task, T-001, asks each architecture to classify a static binary pattern presented once. No memory required — just a single feedforward decision.
Real data — T-001 confirmed results. All approach ceiling.
Almost every architecture approaches perfect accuracy. Dense networks, random graphs, MLP, LSTM, and the MaleCNS-derived model all converge near the ceiling. This validates that our setup works, but it does not reveal whether biological wiring provides any unique advantage. T-001 is too easy.
So we made time matter.
T-002 introduces a memory gap. The model sees a stimulus, then must ignore several intervening steps, and only then recall which pattern it saw at the beginning. You cannot solve this with a single feedforward pass — you must carry information across time.
This mirrors a simple working-memory test: see something, wait, recall it. The gap length is configurable — longer gaps require more persistent memory.
An LSTM has explicit gating mechanisms designed for this. The connectome-derived model relies on its recurrent connectivity and explicit leaky-decay temporal dynamics (SA-010 assumptions).
Then something changed.
On T-002, the MaleCNS-derived model with SA-010 temporal dynamics significantly outperforms the LSTM baseline — despite using far fewer trainable parameters.
Real data — T-002 confirmation (5 seeds · held-out test set).
Not every test has a winner.
The goal of NeuroWeave is not to prove that biological brains are universally better. It is to discover precisely where — and under what conditions — biological structure may provide useful computational properties.
Now explore it.
You've seen the full story — from a real connectome to a controlled benchmark result. Explore the actual numbers, or run the T-002 demonstration yourself.