NeuroWeave
Methodology

From wiring
to benchmark

NeuroWeave treats connectome topology as an architectural prior and asks whether that prior provides measurable computational advantage under controlled tasks.

ABB models are computational. Nothing here recreates a living animal, a complete nervous system, or consciousness.

01

Pipeline

BIOLOGICAL CONNECTOMEMALECNS V1.0 · DROSOPHILA CNSMALECNS-DERIVED ARCHITECTURECONNECTOME-TOPOLOGY-CONSTRAINEDCONTROLLED BENCHMARKT-001 · T-002 · MULTI-SEEDBIOLOGICAL TOPOLOGYCONVENTIONAL AI MODELS

A prototype subgraph is extracted from the connectome and used as a fixed connectivity mask. The same task, data conditions and seed protocol are then applied to the connectome-derived architecture, to topology controls, and to conventional artificial baselines.

02

Connectome source

Dataset

MaleCNS v1.0 — publicly released connectome of the adult male Drosophila central nervous system.

Derived artefact

A 150-node directed prototype subgraph, used as a binary connectivity mask over a trainable weight matrix. Explore it in the connectome explorer.

03

Model families

Bio-constrained
  • A1-BIOMaleCNS-derived architecture: connectivity mask taken from the connectome subgraph, weights trained.
  • A1-FROZENSame MaleCNS-derived topology with the connectome-constrained layer held frozen.
  • A1-BIASMaleCNS-derived topology with bias terms enabled.
  • A1-EDGEMaleCNS-derived topology with edge-weight parameterisation.
Topology controls
  • A3-ERErdős–Rényi random graph control matched on node and edge count.
  • A3-CONFIGConfiguration-model control preserving the degree sequence.
  • A3-DENSEFully connected control over the same node set.
Conventional AI
  • A7-MLPConventional multilayer perceptron baseline.
  • A8-LSTMConventional recurrent baseline (LSTM).
Chance reference
  • A0 RandomUntrained random-output reference establishing chance level.
04

Tasks

05

Evaluation

Each architecture is trained and evaluated independently per random seed under identical data conditions. The reported value is the mean accuracy; the reported spread is the standard deviation across seeds.

06

Limitations

  • The subgraph is a 150-node prototype. It is not the complete Drosophila nervous system and results should not be extrapolated to whole-connectome scale.
  • Nothing in this project recreates a living animal, biological function, or consciousness. All models are computational.
  • T-001 saturates for several architectures, so it cannot discriminate between them.
  • T-002 spreads are wide relative to the differences between conditions; the strongest and weakest seeds differ substantially for some architectures.
  • The SA-010 temporal configuration was selected through exploratory analysis on T-002 and has not been validated as a general setting.
  • Conventional baselines were evaluated at their defaults without architecture-specific tuning.
  • Parameter counts, training times and seed counts are not yet exported by the benchmark backend.