Research · 2026 preprint
Self-Stabilizing Neural Networks via Energy-Guided Structural Adaptation
SSNC lets a network change its own width during training: utility scores guide when neurons are duplicated or removed instead of fixing the architecture upfront. The preprint studies these changes through a regularized energy objective; its CIFAR-10 experiments report bounded neuron counts and decreasing energy, alongside more variable accuracy than fixed-width baselines.
The repository describes an SSNC implementation on TREC-6; the preprint’s abstract reports CIFAR-10 experiments.