Dynamic Continual Learning: Harnessing Parameter Uncertainty for Improved Network Adaptation

Abstract

When fine-tuning Deep Neural Networks (DNNs) to new data, DNNs are prone to overwriting network parameters required for task-specific functionality on previously learned tasks, resulting in a loss of performance on those tasks. We propose using parameter-based uncertainty to determine which parameters are relevant to a network's learned function and regularize training to prevent change in these important parameters. We approach this regularization in two ways: (1), we constrain critical parameters from significant changes by associating more critical parameters with lower learning rates, thereby limiting alterations in those parameters; (2), important parameters are restricted from change by imposing a higher regularization weighting, causing parameters to revert to their states prior to the learning of subsequent tasks. We leverage a Bayesian Moment Propagation framework which learns network parameters concurrently with their associated uncertainties while allowing each parameter to contribute uncertainty to the network's predictive distribution, avoiding the pitfalls of existing sampling-based methods. The proposed approach is evaluated for common sequential benchmark datasets and compared to existing published approaches from the Continual Learning community. Ultimately, we show improved Continual Learning performance for Average Test Accuracy and Backward Transfer metrics compared to sampling-based methods and other non-uncertainty-based approaches.