Brittleness_Plasticity_MEO

Abstract

We introduce Mask Evolution Operators (MEOs), activation-space mechanisms designed to stabilize neural representations during continual learning by applying lightweight restoring forces. MEOs address the fundamental stability–plasticity dilemma by controlling drift at the feature level rather than the weight level. This version clarifies the limitations of earlier pre- liminary experiments. Reported finetune and EWC accuracies (≈6.2%) were obtained from smoke-test runs on an Apple M1 (MPS backend) with drastically shortened training sched- ules. These should be interpreted strictly as diagnostic checks, not benchmarks. Our prior v2 experiments, run with full training schedules on GPUs, achieved substantial improvements (51.2% → 69.1%), supporting the validity of the MEO approach. We formalize two opera- tor variants: Identity, which freezes anchors as a stress-test baseline, and EMA, which allows controlled evolution of class prototypes with per-feature normalization. Identity illustrates rigid- ity, while EMA demonstrates a practical balance between stability and plasticity. We outline a family of extensions—including low-rank subspace anchoring, adaptive stiffness, and hybrid MEO+EWC—that form the basis of ongoing work in Papers 3–5 of the FIL series. This paper also reflects a novel collaborative workflow: the primary draft was co-developed with ChatGPT- 5, with critical feedback from Claude, Grok, and Gemini. By openly documenting iterative testing—including failures and revisions—this series illustrates how human–AI collaboration can accelerate scientific discovery while maintaining transparency.

Comments