A New Regularization-Based Continual Learning Framework

Neural networks outperform all machine-learning models in image classification, generation, and translation. One weakness of such models is that, unlike humans, they cannot learn multiple tasks sequentially. Our study suggests a workable way to train these models sequentially by safeguarding the weights crucial for earlier jobs. This paper proposes a new regularization algorithm for lifelong learning. The algorithm is based on factorized rotation, elastic weight consolidation, and principal component analysis. First, network reparameterization is incorporated by factorized rotation of parameter space. Then, the Fisher information matrix used in elastic weight consolidation identifies important directions in parameter space for each task. After that, Principal Component Analysis (PCA) is used to reduce the dimensions of the Fisher information matrix (FIM) by selectively regularizing weight updates only in unimportant directions for previously learned tasks. Extensive experimentation shows that this proposed algorithm outperforms existing regularization methods for continual learning. Our proposed algorithm performs very well in continuous learning for classification tasks on the MNIST dataset. it is 1% higher than Synaptic Intelligence(SI), 2% higher than best performing regularization-based algorithm Elastic Weight Consolidation(EWC), 1% higher than REWC(Rotated EWC) and 3% higher than Memory Aware-Synapses(MAS).