Efficient Continual Learning in Neural Networks with Embedding Regularization