General Continual Learning through Implicit Contrastive Replays

Abstract

Continual learning in artificial neural networks remains challenging due to catastrophic forgetting, where previously acquired knowledge is rapidly overwritten when learning new tasks. A common approach to mitigate this issue is generative replay, in which a separate generative model is used to approximate past data distributions. However, training such models in a continual learning setting can be unstable and computationally demanding. In this work, we investigate Deep Feedback Control (DFC) as a unified framework for both learning and sample generation. By inverting the network dynamics through a feedback-driven control signal, DFC enables the generation of class-conditional samples without requiring a separate generative model. We analyze the properties of this generation process under different initialization strategies and evaluate the quality of the resulting samples using both qualitative methods and quantitative metrics such as the Fréchet Inception Distance. Building on this mechanism, we propose an internal replay approach for continual learning, in which the same network used for classification also generates replay samples. Experimental results on the SplitMNIST benchmark demonstrate that this method effectively mitigates catastrophic forgetting. In contrast to conventional generative replay, which can suffer from a progressive degradation of sample quality and eventual collapse, DFC-based internal replay exhibits more stable behavior, maintaining task-relevant structure in generated samples over time. Overall, the results suggest that coupling generative and discriminative processes within a single model provides a robust and efficient alternative to traditional replay methods, offering a promising direction for continual learning systems.