Skip to content

NeckLink v3.0.0 — BANC v888 Connectome Weights & Audited Benchmarks (Research Beta)

Choose a tag to compare

@indian-aman indian-aman released this 14 Nov 10:24
· 10 commits to main since this release

Research beta. NeckLink v3.0.0 compiles ~33,000 weight entries from the 157,789-neuron, 23.55M-synapse adult Drosophila BANC v888 connectome into Python (data/*.npz), zero-float/zero-heap ANSI C99 (*_banc_weights.h, 24,024 B Whole-CNS state, 0 B .bss), ESP32/Arduino, and standalone WebAssembly (107.5 KB necklink_cns.wasm), and tests every module against K = 100 Maslov–Sneppen degree- and weight-multiset-preserving raw-neuron shuffles (N = 30 seeds).

NLTP (pynltp==1.6.0) is unchanged in this release.

Highlights & What Was Measured

  • Connectome-derived default weights (connectome_weights="banc_v888") across neckcompass, necksight, necksense, neckcord, and necklink-cns, with the earlier hand-tuned symmetric model retained as connectome_weights="v2_analytic".
  • Connectome extraction & Maslov–Sneppen shuffling (necklink-cns[connectome]): necklink_cns.connectome.extract and necklink_cns.connectome.shuffle plus the reproducible suite and raw outputs in benchmarks/v3/.
  • Audited K = 100 shuffle-gate summary:
    • Where specific connectome topology matters: NeckCompass wave + 125° magnetic spike RMS (12.84° vs 62.82° shuffled, r = 0/100, p = 0.0099), NeckCompass closed-loop homing (100% vs 4.7% shuffled, r = 0/100, p = 0.0099), and NeckCord emergent tripod gait order (0.855 vs 0.411 shuffled, r = 0/100, p = 0.0099); NeckLink CNS closed-loop boat sim (66.7% over 30 seeds; 58.3% vs 4.6% in the 12-seed shuffle gate, r = 2/100, p = 0.030, marginal) and NeckSight thin poles/wires (0.755 vs 0.672 shuffled, r = 3/100, p = 0.040, marginal).
    • Honest nulls: NeckSense mushroom body (0.666 vs 0.672 shuffled on UCI gas novelty, p = 0.74) performs identically to degree-matched random claws.
    • Accuracy vs baselines: On every task, either connectome_weights="v2_analytic" or a tuned classical baseline (χ²-gated EKF, gyro-derotated Farneback, One-Class SVM, nearest centroid, SciPy bounded QP) achieves the highest raw accuracy. Real circuits carry biological left/right asymmetries (e.g., compass drift in darkness).
  • Audit & packaging fixes:
    • Fixed C99/WASM nl_cns_init to initialize plastic_ring.comp with BANC v888 weights (nc_banc_default_weights()).
    • Guarded WASM singletons behind #if defined(__EMSCRIPTEN__) || defined(__wasm__) || defined(NL_CNS_WASM_SINGLETON) so MCU .bss is 0 B (ESP32 esp32dev PlatformIO compile check: 14.5% RAM / 39.3% Flash for NeckLinkCNS).
    • Lazy-loaded scipy/pandas/pyarrow into optional [connectome] extra so runtime wheel installs require only numpy.
    • Added explicit PROVENANCE = "engineering" / "connectome" class labels.
  • Scope & caveats: Pipeline design used real-wiring feedback prior to freezing shuffles; p-values are unadjusted across multiple endpoints; benchmarks other than UCI gas drift and CWRU bearing vibration are numerical/MuJoCo simulations, and closed-loop physical robot hardware has not yet been tested.

PyPI Packages (v3.0.0)

pip install --upgrade necklink==3.0.0
# or individual packages:
pip install --upgrade necksight==3.0.0 neckcompass==3.0.0 neckcord==3.0.0 necksense==3.0.0 necklink-cns==3.0.0