Variance-reduced Language Pretraining via a Mask Proposal Network
This addresses optimization efficiency in large-scale language pretraining, offering an incremental improvement over existing methods.
The paper tackles the slow convergence of self-supervised language pretraining by reducing gradient variance from mask sampling, using a Mask Proposal Network to sample masks based on gradient norms, resulting in faster convergence and higher performance than BERT.
Self-supervised learning, a.k.a., pretraining, is important in natural language processing. Most of the pretraining methods first randomly mask some positions in a sentence and then train a model to recover the tokens at the masked positions. In such a way, the model can be trained without human labeling, and the massive data can be used with billion parameters. Therefore, the optimization efficiency becomes critical. In this paper, we tackle the problem from the view of gradient variance reduction. In particular, we first propose a principled gradient variance decomposition theorem, which shows that the variance of the stochastic gradient of the language pretraining can be naturally decomposed into two terms: the variance that arises from the sample of data in a batch, and the variance that arises from the sampling of the mask. The second term is the key difference between selfsupervised learning and supervised learning, which makes the pretraining slower. In order to reduce the variance of the second part, we leverage the importance sampling strategy, which aims at sampling the masks according to a proposal distribution instead of the uniform distribution. It can be shown that if the proposal distribution is proportional to the gradient norm, the variance of the sampling is reduced. To improve efficiency, we introduced a MAsk Proposal Network (MAPNet), which approximates the optimal mask proposal distribution and is trained end-to-end along with the model. According to the experimental result, our model converges much faster and achieves higher performance than the baseline BERT model.