Online task-free continual learning via discrepancy mechanism

Task Free Continual Learning (TFCL) involves training a deep neural network in a dynamic changing environment defined by unpredictable probabilistic data representation changes. Catastrophic forgetting, which occurs when the network’s weights are replaced following training, is the main factor of performance degeneration in the TFCL. We develop a theoretical framework that accounts for the forgetting process in a continual learning model by deriving the generalization bounds when learning new data while preserving the previously learnt data representations. The theoretical analysis indicates that by dynamically creating new trainable submodels when new information becomes available, can address the challenges of catastrophic forgetting. We then propose the Online Discrepancy Distance Learning (ODDL) model, which expands model’s architecture by evaluating the difference between what was learned by the components of a mixture model and a memory buffer storing the newly available data for training. We then develop a sample selection approach based on a proposed discrepancy distance, which stores only those samples deemed critical to the learning of the model, ensuring the learning of diverse information. The proposed methodology outperforms other static and dynamic expansion models in various TFCL applications.