Year: 2026
Type: article
Abstract: Federated learning (FL) enables collaborative training without centralizing raw data, but remains notoriously challenged by severe client heterogeneity and catastrophic forgetting under streaming distribution shifts. We propose a federated many-to-one (M2O) associative-memory framework designed for the unsupervised recovery of shared archetypes in continual, non-i.i.d. regimes. Departing from standard gradient-exchange methods, each client compresses its local experience into a low-rank Hebbian correlation operator. A central server then aggregates and factorizes these operators into global archetypes via an L-layer Associative Memory (LAM) architecture, strictly preserving raw-data locality. We rigorously cast the server-side aggregation as a low-rank-plus-noise spectral inference problem, utilizing random matrix theory to derive BBP phase transitions for archetype detectability and retrieval robustness: exact at initialization and, in the feedback-driven continual phase, extended through a conditional decomposition of the aggregated operator that leaves the detection threshold unchanged up to an explicit, server-observable inflation factor. To manage the distributed stability-plasticity dilemma, we introduce a local, label-free entropy controller that adaptively modulates client updates. Finally, we show that our adaptive framework, extended to federated Boolean-Gaussian Restricted Boltzmann Machines (RBMs), significantly outperforms standard FedAvg on structured non-IID datasets, validating the spectral mechanism of federated consolidation.
Source: Neural Networks
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $3,110
Awards: G4508289328, G6927059894, G8646124735
OpenAlex ID: W7140130716