ColonelAdaptAI a prototype guided perturbation learning framework for stable continual neural network adaptation

Abstract

Continual learning has become a necessity for intelligent systems operating in dynamic environments where new tasks arrive one after another, such as in multi-task reinforcement learning. While traditional deep learning models perform exceptionally well in static settings, they suffer from catastrophic forgetting of prior knowledge when faced with non-stationary task streams. Previous approaches, including regularisation-based methods, knowledge distillation, and replay-based methods, help alleviate this issue but cannot fully address it. However, regularisation fails on high-visual and semantically similar tasks, distillation can diverge during nonlinear feature adaptation, and replay suffers from excessive memory and computational overhead. This necessitates a need for a solution that balances plasticity, the ability to learn new tasks, with stability, the ability to hold learnt representations, making it clear that these approaches have limitations. To overcome these challenges, we propose ColonelAdaptAI, a continual learning framework based on the Informative Perturbation Network (InfoPerturbNet). The backbone algorithm incorporates task-aware perturbation modulation, mutual–information–guided perturbation refinement, and a lightweight constraint mechanism for stability regularisation. Instead of maintaining large replay buffers or applying weight-anchoring strategies, our framework maintains a discriminative structure via small, semantically aligned perturbation injections at deeper network layers to adapt the model to new tasks. Experimental results on Split MNIST, Permuted MNIST, and Split CIFAR-100 show that the proposed method consistently outperforms the state-of-the-art baselines, improving test accuracy by 4–7% and reducing forgetting by 50% across a long sequence of tasks. Further stability of the coherent evolution of representations is established through perturbation analysis and t-SNE visualisations. This framework offers a pragmatic, efficient, and interpretable approach to continual learning for applications ranging from robotics and health care diagnosis to autonomous systems and any field that requires perpetual learning and generalisation.