Geometric Modulation of Synaptic Plasticity via MGD Metrics for Continual Learning

Read full text Summarise with VIRAT
Authors: diego russo
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Cortical-MGD is a biologically-inspired spiking neural network architecture for continual learning, grounded in the Generative Mathematics of Dimensions (MGD) framework. The core problem is catastrophic forgetting: when a network learns new tasks sequentially, it tends to overwrite previously acquired knowledge. Cortical-MGD addresses this using three geometric quantities computed directly from the weight matrices — Forman-Ricci curvature, SVD effective dimensionality, and the Ryu-Takayanagi min-cut entropy proxy as local signals that regulate synaptic plasticity without any task-boundary supervision. The geometrically modulated learning rate η_mgd scales STDP updates based on the local information geometry of each synapse. Regions with negative curvature (hyperbolic, underexplored) receive higher plasticity for new learning; regions with high effective dimensionality are suppressed to protect existing representations. This modulation alone produces a 3x improvement in forward transfer on sequential MNIST tasks (0.867 vs. 0.283 for a pure STDP baseline) without any architectural change. Three mechanisms complement η_mgd. A modular gate uses the D_avg criterion to select geometrically unspecialized neurons for each new task; switching to random selection increases forgetting by roughly 30x (0.770 vs. 0.025). An area routing rule based on S_RT directs each task to the cortical area with the most residual capacity, without hand-coded thresholds. Offline sleep consolidation replays stored memories in descending S_RT order, analogous to hippocampal slow-wave replay, refining the readout without disturbing consolidated weights. Together these form BioMGDBrain (System G), whose Task-1 retention across five sequential tasks follows a profile of 0.300, 0.917, 1.000, 0.933, 1.000 transient interference at Task 4 resolved through offline replay. Eleven architectures (Systems A through J) are benchmarked on a five-task protocol covering digit classification, geometric patterns, and neuromorphic event data (N-MNIST). System I (Overlap MGD) incorporates 15 biologically motivated mechanisms including neuromodulator-gated plasticity, WTA sparsity control, and synaptic scaling, reaching 1.000 backward retention at both Task 4 and Task 5 using shared cortical areas rather than dedicated ones. System J replaces the Adam readout with a fully biological classifier based on Oja's rule, BCM metaplasticity, and dopamine gating, eliminating backpropagation at every stage. System H (Local MGD) replaces global SVD with per-synapse Forman-Ricci curvature, reducing memory to O(1) per synapse and making the architecture compatible with spike-event neuromorphic hardware such as Intel Loihi. The same geometric framework extends to persistent semantic memory for language models. A text-to-SDR pipeline converts sentence-BERT embeddings into sparse spike patterns that feed a causal cluster routing mechanism, assigning structural identifiers to factual records. Knowledge updates are handled through targeted LTD rather than explicit deletion, so the system can overwrite stale facts without corrupting unrelated memories. On a needle-in-haystack benchmark with 153 records and 50 structurally identical distractors per category, Cortical-MGD achieves 100% Precision@1 vs. 33% for standard vector RAG, and 100% vs. 0% on knowledge-update queries. On a dedicated update benchmark of 10 personal facts with 6 subsequently revised, RAG returns superseded records 50% of the time; Cortical-MGD returns none. The logistic matter field M^σ of the MGD framework acts as a session-context binding signal: on a multi-session benchmark of 50 memories across 5 topics with 15 context-aware queries, Cortical-MGD achieves 80.0% Precision@1 vs. 73.3% for cosine-only retrieval.

Fields of Study & Subjects

Domain Adaptation and Few-Shot Learning Neural Networks and Applications Advanced Memory and Neural Computing

Publication details

Journal / source
Zenodo (CERN European Organization for Nuclear Research)
Publication date
25 March 2026
Language
EN