Learning to Learn without Forgetting by Maximizing Transfer and Minimizing Interference

Lack of performance when it comes to continual learning over non-stationary of data remains a major challenge in scaling neural network to more human realistic settings. In this work we propose a new of the continual learning problem in terms of a temporally trade-off between transfer and interference that can be optimized by gradient alignment across examples. We then propose a new algorithm, -Experience Replay (MER), that directly exploits this view by combining replay with optimization based meta-learning. This method learns that make interference based on future gradients less likely and based on future gradients more likely. We conduct experiments across lifelong supervised learning benchmarks and non-stationary learning environments demonstrating that our approach outperforms recently proposed baselines for continual learning. experiments show that the gap between the performance of MER and baseline grows both as the environment gets more non-stationary and as the of the total experiences stored gets smaller.