Non-Equilibrium Stochastic Dynamics as a Unified Framework for Insight and Repetitive Learning: A Kramers Escape Approach to Continual Learning

Gunn Kim

cond-mat.stat-mechApr 5, 2026

Non-Equilibrium Stochastic Dynamics as a Unified Framework for Insight and Repetitive Learning: A Kramers Escape Approach to Continual Learning

Authors: Gunn Kim

Organizations: Department of Physics, Sejong University, Seoul 05006, Republic of Korea

Abstract

Continual learning in artificial neural networks is fundamentally limited by the stability--plasticity dilemma: systems that retain prior knowledge tend to resist acquiring new knowledge, and vice versa. Existing approaches, most notably elastic weight consolidation~(EWC), address this empirically without a physical account of why plasticity eventually collapses as tasks accumulate. Separately, the distinction between sudden insight and gradual skill acquisition through repetitive practice has lacked a unified theoretical description. Here, we show that both problems admit a common resolution within non-equilibrium statistical physics. We model the state of a learning system as a particle evolving under Langevin dynamics on a double-well energy landscape, with the noise amplitude governed by a time-dependent effective temperature $T(t)$ . The probability density obeys a Fokker--Planck equation, and transitions between metastable states are governed by the Kramers escape rate $k = (ω_0ω_b/2π)\,e^{-ΔE/T}$ . We make two contributions. First, we identify the EWC penalty term as an energy barrier whose height grows linearly with the number of accumulated tasks, yielding an exponential collapse of the transition rate predicted analytically and confirmed numerically. Second, we show that insight and repetitive learning correspond to two qualitatively distinct temperature protocols within the same Fokker--Planck equation: insight events produce transient spikes in $T(t)$ that drive rapid barrier crossing, whereas repetitive practice operates at a modestly elevated but fixed temperature, achieving transitions through sustained stochastic diffusion. These results establish a physically grounded framework for understanding plasticity and its failure in continual learning systems, and suggest principled design criteria for adaptive noise schedules in artificial intelligence.

Figures & tables

Figure 1: Model setup. (a) Double-well energy landscape

E(s)=(s^{2}-1)^{2}

with the three temperature levels indicated. The barrier height is

\Delta E=1

, with minima at

s=\pm 1

and barrier at

s=0

. (b) Schematic of the three temperature protocols: fixed

T_{0}=0.22

(EWC-like, blue dashed), repetitive training at elevated fixed

T_{R}=0.32

(green dash-dot), and adaptive

T(t)

with transient insight spikes to

T_{\rm kick}=0.95

(red solid). The equally spaced spikes are illustrative; in the simulation, spikes occur periodically with interval

\Delta t_{\rm kick}=50\,\rm s

as a simplified proxy for event-driven triggers such as prediction error or novelty signals.

Figure 2: Simulation results for the three protocols ( $T_{0}=0.22$ , $T_{R}=0.32$ , $T_{\rm kick}=0.95$ ; $N=6\times 10^{5}$ steps, $\Delta t=10^{-3}$ ). (a)–(c) Representative trajectories showing zero transitions under fixed $T_{0}$ , frequent transitions under adaptive $T(t)$ , and occasional transitions under repetitive $T_{R}$ . The number of well-to-well transitions is shown in each panel. (d) Time series of $T(t)$ for the adaptive protocol; vertical lines mark transition events. (e) Steady-state probability density $\rho(s)$ for all three protocols, compared to the Boltzmann distribution at $T_{0}$ (black dotted). The adaptive protocol produces a near-symmetric bimodal distribution; the fixed protocol remains unimodal. The shaded region indicates the density gain of the adaptive protocol relative to the fixed protocol.

Figure 3: Quantitative validation against Kramers theory. (a) Transition rates measured in simulation (well-to-well criterion,

|s|>0.7

) across a sweep of fixed temperatures on a logarithmic

y

-axis, compared to the Kramers curve

k=(\omega_{0}\omega_{b}/2\pi)\,e^{-\Delta E/T}

(solid black line). The linear relationship between

\log k

and

1/T

(Arrhenius form) is clearly visible. Gray circles: fixed-

T

sweep. Colored markers: the three operating points (fixed

T_{0}

, blue square; adaptive

\langle T\rangle

, red triangle; repetitive

T_{R}

, green diamond). (b) Transition rates for the three protocols; dashed lines indicate Kramers predictions.

Figure 4: EWC plasticity collapse. (a) Transition rate as a function of the number of accumulated tasks

n

for EWC (theory, blue dashed line; simulation, open squares) and the adaptive

T(t)

protocol (red triangles). The EWC rate collapses exponentially as predicted by Eq. ( 10 ), while the adaptive protocol maintains a constant rate. The repetitive protocol (green diamonds) provides a constant but uncompensated reference. (b) Barrier height

\Delta E(n)

(black squares, left axis) and the temperature

T(n)

required to maintain a constant transition rate (red triangles, right axis) as functions of

n

Continual Learning

Continual Learning studies how models acquire knowledge from changing data streams while retaining or selectively revising what they learned before.

See all 304 papers

Jun 1, 2026

Continual Learning as a Multiphase Moving-Boundary Problem

Continual learning struggles to balance retaining past knowledge with absorbing new tasks. Stefan-CL elegantly resolves this stability-plasticity dilemma through the physics of melting. It frames consolidated knowledge as a protected "solid" and unused capacity as an adaptable "liquid." As the network learns, this boundary expands, governed by a "latent heat" tuning dial. By mathematically freezing the learned interior, Stefan-CL cuts forgetting to near zero, matching memory-heavy baselines without storing raw data, forging a beautiful, physics-grounded path for AI.

Snigdha Chandan Khilar

Independent Researcher

Continual Learning Catastrophic Forgetting

Ask PDF

Jun 8, 2026

Preserving Plasticity in Continual Learning via Dynamical Isometry

Continual training of deep neural networks under non-stationarity often leads to a progressive loss of plasticity, eventually limiting further learning. We relate plasticity to the empirical Neural Tangent Kernel, and identify dynamical isometry (the condition that layer-wise Jacobian singular values remain close to one) as a key mechanism for preserving plasticity in continual learning. We revisit a class of networks that are almost-everywhere isometric while remaining universal Lipschitz function approximators, demonstrating that near-dynamical isometry is compatible with expressive nonlinear representations. For general architectures, we propose an efficient isometry-promoting regularization scheme and identify a novel mechanism by which it can reactivate dormant ReLU units. Building on this, we introduce AdamO, an Adam-style adaptive optimizer that decouples isometry regularization from gradient updates, analogous to AdamW. We further reinterpret prior plasticity-preserving approaches through the lens of dynamical isometry, showing that they target only a partial measure of isometry. Across supervised and reinforcement-learning continual-learning benchmarks designed to induce plasticity loss, our methods consistently match or outperform existing approaches.

Andries Rosseau, Robert Müller, Ann Nowé

AI Lab, Vrije Universiteit Brussel · Aganthos

Loss of Plasticity Continual Learning Adaptive Gradient Methods Neural Network Training Dynamics

Ask PDF

Oct 27, 2026

Beyond Parameter Protection: Fisher-Shaped EWC for Replay-Free Continual Learning

In replay-free continual learning, Elastic Weight Consolidation (EWC) uses Fisher information matrix (FIM) to protect acquired knowledge by penalizing changes to historically important parameters. However, a local importance estimate at the retained solution does not in general characterize sensitivity after finite parameter displacement. This motivates controlling Fisher sensitivity over a neighborhood of potential future updates. We propose Fisher-Shaped EWC (FS-EWC), which extends the role of FIM from parameter protection to solution formation for future retention. During task learning, FS-EWC minimizes the worst-case empirical FIM trace within a prescribed neighborhood, explicitly favoring solutions that are less sensitive to bounded future displacement. At task completion, trajectory-based model merging constructs the retained solution from multiple optimization states rather than a single terminal iterate, after which FIM is estimated for subsequent consolidation. Experiments across multiple benchmarks show that FS-EWC improves performance over EWC, while controlled studies indicate a more favorable stability--plasticity balance and greater tolerance to parameter drift. These results suggest that effective knowledge retention requires both forming solutions with future retention in mind and protecting acquired knowledge.

Liying Wang, Simone Magistri, Joost van de Weijer

Department of Computer Science, Universitat Aut`onoma de Barcelona, Spain · Computer Vision Center, Barcelona, Spain · Media Integration and Communication Center, University of Florence, Italy

Elastic Weight Consolidation Regularization-Based Continual Learning Catastrophic Forgetting Fisher Information

Ask PDF

Abstract

Continual Learning

Explore similar work

Continual Learning as a Multiphase Moving-Boundary Problem

Preserving Plasticity in Continual Learning via Dynamical Isometry

Beyond Parameter Protection: Fisher-Shaped EWC for Replay-Free Continual Learning