Hebbian-Oscillatory Co-Learning

Hasi Hays · 2026

Read Hebbian-Oscillatory Co-Learning by Hasi Hays (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-08731.txt.

Introduction

Hasi Hays unites two brain-like learning ideas in one two-timescale framework. Sparse geometric nets give Hebbian-driven shape, while oscillator attention swaps dot products for phase locking. A global order signal gates structural updates, so links firm up only when phase coherence marks a real pattern, and the joint system provably settles to a stable rest state.

Catalog record

CREATORHasi Hays
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE64793
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.08731

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Hebbian-Oscillatory Co-Learning Hasi Hays Abstract We introduce Hebbian-Oscillatory Co-Learning (HOC-L), a unified two-timescale dynamical framework for joint structural plasticity and phase synchronization in bio-inspired sparse neural architectures. HOC-L couples our two recent frameworks: the hyperbolic sparse geometry of Resonant Sparse Geometry Networks (RSGN), which employs Poincaré ball embeddings with Hebbian-driven dynamic sparsity, and the oscillator-based attention of Selective Synchronization Attention (SSA), which replaces dot-product attention with Kuramoto-type phase-locking dynamics. The key mechanism is synchronization-gated plasticity: the macroscopic order parameter r ⁡ ( t ) r(t) of the oscillator ensemble gates Hebbian structural updates, so that connectivity consolidation occurs only when sufficient phase coherence signals a meaningful computational pattern. We prove convergence of the joint system to a stable equilibrium via a composite Lyapunov function and derive explicit timescale separation bounds. The resulting architecture achieves O ⁡ ( n ⋅ k ) O(n\cdot k) complexity with k ≪ n k\ll n , preserving the sparsity of both parent frameworks. Numerical simulations confirm the theoretical predictions, demonstrating emergent cluster-aligned connectivity and monotonic Lyapunov decrease. I Introduction The capacity of biological neural systems to learn, adapt, and compute efficiently arises from a sophisticated interplay between multiple mechanisms operating on different temporal scales 7 . At the slow timescale of hours to days, synaptic connections strengthen or weaken according to Hebbian plasticity rules: neurons that fire together wire together 22 . At the fast timescale of milliseconds, oscillatory synchronization among neural populations coordinates information flow and gates communication between brain regions 14 ; 33 . These two processes are not independent; rather, accumulating neuroscientific evidence indicates that oscillatory coherence directly modulates the conditions under which structural plasticity occurs 11 ; 28 . Phase-locked neural assemblies preferentially strengthen their mutual synaptic connections, while desynchronized populations undergo synaptic weakening. Despite this well-established biological principle, contemporary artificial neural network architectures treat structural learning and dynamic coordination as separate computational concerns. Standard deep learning 20 optimizes a fixed network topology through gradient descent, with no mechanism for activity-dependent structural rewiring. Sparse network methods such as the lottery ticket hypothesis 13 and adaptive sparse training 29 address the structural dimension but lack any oscillatory coordination mechanism. Conversely, oscillator-based neural computation models 27 ; 23 capture temporal coordination but typically operate on fixed connectivity graphs without plasticity. In our recent work, we have developed two complementary frameworks that each address one side of this gap. Resonant Sparse Geometry Networks (RSGN) 18 introduced a brain-inspired architecture that embeds neural representations in the Poincaré ball model of hyperbolic space, enabling naturally hierarchical and sparse connectivity patterns. Crucially, RSGN employs a two-timescale learning scheme in which fast gradient-based optimization handles immediate task objectives, while slow Hebbian updates reshape the structural connectivity of the sparse graph. This input-dependent dynamic sparsity yields architectures where the active computational subnetwork adapts to each input, achieving substantial efficiency gains without sacrificing expressiveness. In a companion study, we proposed Selective Synchronization Attention (SSA) 21 , which replaces the standard dot-product self-attention mechanism 19 in transformers with a Kuramoto oscillator model. In SSA, each token is assigned a learnable intrinsic frequency, and attention weights emerge from the degree of phase-locking between token oscillators rather than from inner products in key-query space. The macroscopic order parameter r(t) of the oscillator ensemble provides a natural, differentiable sparsity mechanism: only sufficiently synchronized token pairs receive non-negligible attention weight, while desynchronized pairs are effectively masked. Together, these two frameworks address complementary aspects of the same fundamental biological phenomenon. RSGN captures the structural plasticity that operates on the slow timescale, namely how connectivity patterns emerge from correlated activity, while SSA captures the oscillatory dynamics that operate on the fast timescale, namely how phase coherence selects which information flows through existing connections. Their unification is not merely an engineering convenience but a desirable property for any model that aspires to capture the full richness of neural computation. We introduce Hebbian-Oscillatory Co-Learning (HOC-L) , a unified two-timescale dynamical system that formally couples structural plasticity with phase synchronization. The core insight of HOC-L is that the order parameter r(t) from SSA’s oscillator dynamics serves as a natural gate for RSGN’s Hebbian structural updates. When a sufficient subset of oscillators achieves phase coherence ( r(t)>r_{c} for a critical threshold r_{c} ), this signals that the network has identified a meaningful computational pattern, and the corresponding structural connections are strengthened via a Hebbian rule. This creates a virtuous cycle: synchronization identifies important pathways, Hebbian plasticity consolidates them structurally, and the resulting structural changes facilitate future synchronization. The principal contributions of this paper are as follows: 1. We formulate the first unified dynamical system coupling Kuramoto-type phase synchronization with synchronization-gated Hebbian structural plasticity, grounded in the frameworks of RSGN 18 and SSA 21 . 2. We prove convergence of the joint two-timescale system to a stable equilibrium using a composite Lyapunov function, establishing explicit conditions on the timescale separation ratio \tau_{\text{fast}}/\tau_{\text{slow}} . 3. We derive the complete HOC-L architecture with O(n\cdot k) computational complexity where k\ll n , preserving the sparsity benefits of both parent frameworks. 4. We propose comprehensive experimental protocols for validating the theoretical predictions across multiple benchmark domains. We develop the complete theoretical framework, present convergence proofs and stability analysis, describe the concrete architecture with supporting numerical simulations, and outline experimental protocols for empirical validation. II Related Work HOC-L draws upon and extends several established research directions spanning neuroscience-inspired learning rules, oscillatory computation, geometric representations, and sparse architectures. The principle that correlated neural activity drives synaptic strengthening was first articulated by Hebb 22 and subsequently refined through the discovery of spike-timing-dependent plasticity 3 ; 8 . In the machine learning context, Hebbian principles have informed unsupervised feature learning, competitive learning, and self-organizing maps 22 ; 3 ; 28 . In our recent work, RSGN 18 , we demonstrated that Hebbian updates can be integrated into a gradient-trained sparse architecture operating in hyperbolic space, enabling input-dependent structural adaptation at a slow timescale while fast gradient updates handle task-specific optimization. This two-timescale design principle provides one of the two foundational pillars of the present work. The second pillar originates from oscillatory neural computation. The Kuramoto model 27 provides the canonical mathematical framework for synchronization in coupled oscillator populations, and its application to neuroscience has elucidated how gamma-band oscillations coordinate neural communication 14 ; 7 and how phase coherence supports memory formation 11 . More broadly, the integration of continuous dynamical systems into neural architectures has been advanced by Neural ODEs 10 , which parameterize hidden state evolution as ordinary differential equations; HOC-L shares the continuous-time perspective but focuses specifically on oscillatory phase dynamics rather than general-purpose ODE solvers. The theoretical analysis of Kuramoto dynamics, including the characterization of phase transitions via the macroscopic order parameter, has been extensively developed 34 ; 1 ; 6 . In our companion work, SSA 21 , we brought these ideas into the transformer architecture by replacing dot-product attention with Kuramoto-type synchronization weights, using the order parameter r as a coherence measure and a learned compatibility kernel C(\omega_{i},\omega_{j}) to modulate coupling strength. The present work extends SSA’s oscillatory dynamics by coupling them to structural plasticity through a synchronization gate. The geometric substrate for HOC-L’s structural component builds on advances in graph neural networks and hyperbolic

Read the complete text