Apoptotic Weight Transfer in Self-Organizing Maps: A Multi-Generational Framework with Optimal Parameter Regimes for Neuromorphic Resilience

Abstract

AbstractWerevisit Apoptotic Weight Transfer (AWT), a mechanism in which a dying neuron’s learned representation isredistributedtohealthyneighborsbeforequarantine, embeddedinamulti-generationalSelf-OrganizingMap(SOM)withadaptivequarantineandhealing. Earlier drafts reported large improvements over baselines.Those numbers did not survive careful re-examination. Three problems were found: baselines were attackedmore harshly than AWT, the PGD attack was implemented incorrectly, and the attack was applied to only asmall subset of patterns.This paper reports a fair re-evaluation on Synthetic, MNIST, and Fashion-MNIST, with 5 seeds andt-based 95% confidence intervals. Under a fair protocol, AWT shows strong scalability (from 4.52% QEdegradation at 8×8 to 0.10% at 32×32) and a clean failure-mode boundary, but it does not outperformseveral of its own ablations. Disabling the healing mechanism reduces QE degradation from 4.52% to0.00%—healing, which we designed as a strength, is harmful under concentrated attack. The variancescaling model fails on log-scale fitting (𝑅2 ≈ 0.36), and the antibody-like behavior is indistinguishable fromnoise (𝐹1 ≈ 0.38±0.37). We report these results honestly and identify three structural design principles thatany apoptosis-inspired learning system must satisfy.