刘婷

Yiming Cui, Ziqing Yang, Ting Liu · deepmind

Pre-trained Language Models (PLMs) have been widely used in various natural language processing (NLP) tasks, owing to their powerful text representations trained on large-scale corpora. In this paper, we propose a new PLM called PERT for natural language understanding (NLU). PERT is an auto-encoding model (like BERT) trained with Permuted Language Model (PerLM). The formulation of the proposed PerLM is straightforward. We permute a proportion of the input text, and the training objective is to predict the position of the original token. Moreover, we also apply whole word masking and N-gram masking to improve the performance of PERT. We carried out extensive experiments on both Chinese and English NLU benchmarks. The experimental results show that PERT can bring improvements over various comparable baselines on some of the tasks, while others are not. These results indicate that developing more diverse pre-training tasks is possible instead of masked language model variants. Several quantitative studies are carried out to better understand PERT, which might help design PLMs in the future. Resources are available: https://github.com/ymcui/PERT

Juntang Zhuang, Boqing Gong, Liangzhe Yuan et al. · deepmind

The recently proposed Sharpness-Aware Minimization (SAM) improves generalization by minimizing a \textit{perturbed loss} defined as the maximum loss within a neighborhood in the parameter space. However, we show that both sharp and flat minima can have a low perturbed loss, implying that SAM does not always prefer flat minima. Instead, we define a \textit{surrogate gap}, a measure equivalent to the dominant eigenvalue of Hessian at a local minimum when the radius of the neighborhood (to derive the perturbed loss) is small. The surrogate gap is easy to compute and feasible for direct minimization during training. Based on the above observations, we propose Surrogate \textbf{G}ap Guided \textbf{S}harpness-\textbf{A}ware \textbf{M}inimization (GSAM), a novel improvement over SAM with negligible computation overhead. Conceptually, GSAM consists of two steps: 1) a gradient descent like SAM to minimize the perturbed loss, and 2) an \textit{ascent} step in the \textit{orthogonal} direction (after gradient decomposition) to minimize the surrogate gap and yet not affect the perturbed loss. GSAM seeks a region with both small loss (by step 1) and low sharpness (by step 2), giving rise to a model with high generalization capabilities. Theoretically, we show the convergence of GSAM and provably better generalization than SAM. Empirically, GSAM consistently improves generalization (e.g., +3.2\% over SAM and +5.4\% over AdamW on ImageNet top-1 accuracy for ViT-B/32). Code is released at \url{ https://sites.google.com/view/gsam-iclr22/home}.