Explicit Inductive Bias for Transfer Learning with Convolutional\n Networks

In inductive transfer learning, fine-tuning pre-trained convolutional\nnetworks substantially outperforms training from scratch. When using\nfine-tuning, the underlying assumption is that the pre-trained model extracts\ngeneric features, which are at least partially relevant for solving the target\ntask, but would be difficult to extract from the limited amount of data\navailable on the target task. However, besides the initialization with the\npre-trained model and the early stopping, there is no mechanism in fine-tuning\nfor retaining the features learned on the source task. In this paper, we\ninvestigate several regularization schemes that explicitly promote the\nsimilarity of the final solution with the initial model. We show the benefit of\nhaving an explicit inductive bias towards the initial model, and we eventually\nrecommend a simple $L^2$ penalty with the pre-trained model being a reference\nas the baseline of penalty for transfer learning tasks.\n