Continual Learning in Convolutional Neural Networks with Tensor Rank Updates

Continual learning (CL) focuses on training machine learning models to incrementally learn new tasks while avoiding “catastrophic forgetting” of previous tasks. CL is analogous to human learning and effective computational techniques are being actively researched. This paper introduces a novel approach to CL in convolutional neural networks (CNN s) using tensor rank increments, which facilitates learning new tasks without forgetting previous ones. For each new task, our method intelli-gently reuses parameters trained for previous tasks, augmented with a small number of new parameters, thereby achieving efficient learning. Our experimental studies demonstrate the superior performance of the proposed method compared to existing alternatives across multiple standard datasets.