#19 · Systems biology

Local normalization, pruning, and learning rules shape recurrent network computation

Three preprints from August 2026 examine how locally applied constraints shape the computations recurrent networks sustain. RDNN shows that divisive normalization produces stable continuous working-memory slow manifolds with low effective rank. Noise-prune, an unsupervised local pruning rule applied to task-trained recurrent networks, preserves performance under sampling and connection rescaling. The layered surprise cascade variant modifies the Forward-Forward objective in a recurrent setting, reporting predictive, cortex-like representations from local learning rules. The items differ in scale and method, so the shared thread is conceptual: normalization and locally applied rules are treated as variables determining which computations a network can readily perform, rather than connectivity alone.