New theory predicts forgetting and coexistence in continual learning
Neural Succession: A Mesoscopic Theory of Invasion, Coexistence, and Stabilization in Continual Learning
Machine Learning
Summary
Continual learning means training AI models on new tasks without forgetting old ones. The authors introduce Successional Learning Theory, which treats learned knowledge like a community facing an invading task. Their theory measures how compatible an incoming task is with current knowledge to predict when forgetting happens or if old and new knowledge can coexist. They tested this idea extensively across many datasets and found it predicts forgetting better than previous methods. This helps improve understanding of continual learning and how to maintain knowledge over time.
What this means in practice
- •For machine learning engineers: Predict when training on new tasks will cause forgetting to design better continual learning models with built-in diagnostics.
- •For ai system maintainers: Assess compatibility between new and existing knowledge to optimize memory replay strategies and reduce data storage needs.
Authors
Shoaib Ahmed Dipu, Md Salman Shamil, Sayeed Shafayet Chowdhury
Abstract
Continual learning is usually studied through mechanisms that preserve old knowledge. We develop Successional Learning Theory (SLT), a mesoscopic account in which the current representation is a resident community, the incoming task is an invader, forgetting is resident displacement, joint retention is coexistence, replay is resident reinforcement, and training moves from establishment toward stabilization. Its empirical coordinate is directional pre-invasion compatibility, measured on the resident model before the incoming task is learned. Across eight experiments, compatibility orders later forgetting on the 20 directed Split-CIFAR-10 transitions (three-repeat r=-0.789, incoming-task cluster 95% CI [-0.90,-0.72], every repeat alone r<=-0.67), forecasts held-out forgetting with 24% lower error than a no-information baseline, and reproduces under controlled MNIST permutations and CIFAR-10 rotations (r=-0.804, -0.718). On an 84-transition suite, compatibility separates coexistence from exclusion at every retention threshold (AUC 0.93-0.97). Replay repairs every transition with at most 325 stored examples and is most efficient where displacement is largest. Compatibility reaches |r|=0.720, while activation, representation, Jacobian, and fixed-coefficient Lotka-Volterra specializations do not. Plasticity and feature turnover fall reliably from early to late training (15/15 and 14/15 runs). We formalize a minimum habitat-modification bound, a displacement floor, a sufficient coexistence condition, an identifiability law with a range-restriction corollary, successional stabilization, and local reinforcement. The identifiability law also predicts where the coordinate loses leverage, and the prediction matches three CIFAR-100 partitions and five optimizer regimes. SLT is a pre-adaptation diagnostic that complements replay, regularization, and projection methods.