Answers are grounded in this paper and cite their sources.
The study demonstrates a new learning rule for synaptic stability, suggesting improved mechanisms for neural networks.
Hebbian learning — Δw = η · x_pre · x_post — has served as the dominant model of synaptic plasticity for over seven decades. We demonstrate that pure Hebbian learning is provably unstable: under the standard update rule, synaptic weights diverge without external normalization. The field's response has been a succession of stabilization mechanisms — Oja's rule, BCM sliding threshold, homeostatic synaptic scaling, weight normalization — which we argue do not repair Hebbian learning but instead constitute evidence that the update rule is structurally wrong. A correct learning rule must be self-stabilizing by construction, not by external constraint. We derive an alternative from the Scale-Recursive Non-Uniform Domain Tiling (SRNUDT) framework. In SRNUDT, a domain emits remainder when it cannot contain its input. Remainder is not a byproduct — it is the error signal. We show that treating remainder emission as the learning signal yields a rule with three properties that Hebbian learning lacks by construction: (1) negative feedback stability — overflow reduces the weights that caused it; (2) self-termination — learning stops when the substrate contains its input; (3) proportional input attribution — only edges that contributed to the overflow are adjusted, in exact proportion to their delivered signal fraction. (Note: this constitutes per-node attribution, not network-level credit assignment across layers; the distinction is discussed in Section 5.3 and 7.4.) We formalize the rule, prove its stability and self-termination properties, and demonstrate its behavior in RemainderOS, a SRNUDT-native substrate implementation. The hub node problem observed under Hebbian — runaway degree accumulation at a single node while average degree remains bounded — is resolved under remainder propagation learning through the same negative feedback mechanism.
No takes yet. Share an insight, caveat, or question.
Bradley Ploof (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: