partial-annealing-pattern-decorrelation

name

partial-annealing-pattern-decorrelation

description

Partial annealing framework for associative neural networks. Core idea: Coupling neural dynamics with slowly evolving patterns via two-temperature-two-timescale framework introduces real parameter n (replica-like) that tunes fast-slow separation. Negative n induces pattern decorrelation, reducing interference and promoting orthogonality, achieving maximal storage capacity αc=1. Outperforms standard decorrelation on biased patterns. Activation: partial annealing, pattern decorrelation, associative memory, Hopfield model, two-temperature framework, replica trick, memory capacity, pattern interference, synaptic plasticity timescale, neural memory organization.

Partial Annealing and Pattern Decorrelation in Associative Neural Networks

Paper

• Title: Partial annealing and pattern decorrelation in associative neural networks
• Authors: Linda Albanese, Andrea Alessandrelli, Adriano Barra, Silvio Franz, Federico Ricci-Tersenghi
• arXiv: 2605.10304v1 (2026-05-11)
• Category: cond-mat.dis-nn

Core Methodology

Problem

Standard associative networks (Hopfield model) suffer from pattern interference when storing many or biased patterns, limiting storage capacity.

Key Framework: Two-Temperature-Two-Timescale

Couple fast degrees of freedom (neurons) with slow degrees of freedom (synapses/patterns):

• Neurons evolve at temperature T₁, fast timescale
• Patterns evolve at temperature T₂, slow timescale
• Separation parameter n (replica-like, real-valued) tunes the fast-slow gap

Key Findings

1. Negative n → Pattern Decorrelation: Induces progressive decorrelation of stored patterns, reducing interference
2. Maximal Capacity: Achieves theoretical maximum storage αc = 1
3. Biased Patterns: Restores retrieval where standard methods fail
4. Adaptive Mechanism: Partial annealing serves as self-organizing memory optimization

Derivation Technique

• Adapts Guerra's interpolation to the partial annealing case
• Derives free energy without analytical continuation
• Validated via mean-field Monte Carlo simulations

Implications

• Offers principled approach to memory organization in complex systems
• Suggests biological systems may use timescale separation for memory optimization
• Relevant for continual learning: slow synaptic consolidation + fast neural dynamics
• Outperforms explicit decorrelation methods on non-trivial pattern distributions

Usage Triggers

• Analyzing associative memory capacity limits
• Studying pattern interference in Hopfield-type networks
• Designing continual learning systems with timescale separation
• Understanding biological memory consolidation mechanisms
• Optimizing storage in associative neural networks

Related Skills

• attractor-metadynamics-neural: Neural attractor landscape dynamics
• kernel-hopfield-attractor-geometry: Hopfield network attractor analysis
• snn-working-memory-heterogeneous-delays: Working memory in spiking networks