Dong Li

Hongyu Wang, Shuming Ma, Li Dong et al. · microsoft-research

In this paper, we propose a simple yet effective method to stabilize extremely deep Transformers. Specifically, we introduce a new normalization function (DeepNorm) to modify the residual connection in Transformer, accompanying with theoretically derived initialization. In-depth theoretical analysis shows that model updates can be bounded in a stable way. The proposed method combines the best of two worlds, i.e., good performance of Post-LN and stable training of Pre-LN, making DeepNorm a preferred alternative. We successfully scale Transformers up to 1,000 layers (i.e., 2,500 attention and feed-forward network sublayers) without difficulty, which is one order of magnitude deeper than previous deep Transformers. Remarkably, on a multilingual benchmark with 7,482 translation directions, our 200-layer model with 3.2B parameters significantly outperforms the 48-layer state-of-the-art model with 12B parameters by 5 BLEU points, which indicates a promising scaling direction.

Yutao Sun, Li Dong, Barun Patra et al. · microsoft-research, tsinghua

Position modeling plays a critical role in Transformers. In this paper, we focus on length extrapolation, i.e., training on short texts while evaluating longer sequences. We define attention resolution as an indicator of extrapolation. Then we propose two designs to improve the above metric of Transformers. Specifically, we introduce a relative position embedding to explicitly maximize attention resolution. Moreover, we use blockwise causal attention during inference for better resolution. We evaluate different Transformer variants with language modeling. Experimental results show that our model achieves strong performance in both interpolation and extrapolation settings. The code will be available at https://aka.ms/LeX-Transformer.

Hongyu Wang, Shuming Ma, Shaohan Huang et al. · cmu, microsoft-research

A big convergence of model architectures across language, vision, speech, and multimodal is emerging. However, under the same name "Transformers", the above areas use different implementations for better performance, e.g., Post-LayerNorm for BERT, and Pre-LayerNorm for GPT and vision Transformers. We call for the development of Foundation Transformer for true general-purpose modeling, which serves as a go-to architecture for various tasks and modalities with guaranteed training stability. In this work, we introduce a Transformer variant, named Magneto, to fulfill the goal. Specifically, we propose Sub-LayerNorm for good expressivity, and the initialization strategy theoretically derived from DeepNet for stable scaling up. Extensive experiments demonstrate its superior performance and better stability than the de facto Transformer variants designed for various applications, including language modeling (i.e., BERT, and GPT), machine translation, vision pretraining (i.e., BEiT), speech recognition, and multimodal pretraining (i.e., BEiT-3).

Dong Li, Shuisheng Zhou, Witold Pedrycz

Fuzzy C-Means (FCM) is a widely used clustering method. However, FCM and its many accelerated variants have low efficiency in the mid-to-late stage of the clustering process. In this stage, all samples are involved in the update of their non-affinity centers, and the fuzzy membership grades of the most of samples, whose assignment is unchanged, are still updated by calculating the samples-centers distances. All those lead to the algorithms converging slowly. In this paper, a new affinity filtering technique is developed to recognize a complete set of the non-affinity centers for each sample with low computations. Then, a new membership scaling technique is suggested to set the membership grades between each sample and its non-affinity centers to 0 and maintain the fuzzy membership grades for others. By integrating those two techniques, FCM based on new affinity filtering and membership scaling (AMFCM) is proposed to accelerate the whole convergence process of FCM. Many experimental results performed on synthetic and real-world data sets have shown the feasibility and efficiency of the proposed algorithm. Compared with the state-of-the-art algorithms, AMFCM is significantly faster and more effective. For example, AMFCM reduces the number of the iteration of FCM by 80% on average.

Dong Li, Shuisheng Zhou, Tieyong Zeng et al.

K-Means algorithm is a popular clustering method. However, it has two limitations: 1) it gets stuck easily in spurious local minima, and 2) the number of clusters k has to be given a priori. To solve these two issues, a multi-prototypes convex merging based K-Means clustering algorithm (MCKM) is presented. First, based on the structure of the spurious local minima of the K-Means problem, a multi-prototypes sampling (MPS) is designed to select the appropriate number of multi-prototypes for data with arbitrary shapes. A theoretical proof is given to guarantee that the multi-prototypes selected by MPS can achieve a constant factor approximation to the optimal cost of the K-Means problem. Then, a merging technique, called convex merging (CM), merges the multi-prototypes to get a better local minima without k being given a priori. Specifically, CM can obtain the optimal merging and estimate the correct k. By integrating these two techniques with K-Means algorithm, the proposed MCKM is an efficient and explainable clustering algorithm for escaping the undesirable local minima of K-Means problem without given k first. Experimental results performed on synthetic and real-world data sets have verified the effectiveness of the proposed algorithm.

Chen Chen, Hongyao Tang, Yi Ma et al.

Pessimism is of great importance in offline reinforcement learning (RL). One broad category of offline RL algorithms fulfills pessimism by explicit or implicit behavior regularization. However, most of them only consider policy divergence as behavior regularization, ignoring the effect of how the offline state distribution differs with that of the learning policy, which may lead to under-pessimism for some states and over-pessimism for others. Taking account of this problem, we propose a principled algorithmic framework for offline RL, called \emph{State-Aware Proximal Pessimism} (SA-PP). The key idea of SA-PP is leveraging discounted stationary state distribution ratios between the learning policy and the offline dataset to modulate the degree of behavior regularization in a state-wise manner, so that pessimism can be implemented in a more appropriate way. We first provide theoretical justifications on the superiority of SA-PP over previous algorithms, demonstrating that SA-PP produces a lower suboptimality upper bound in a broad range of settings. Furthermore, we propose a new algorithm named \emph{State-Aware Conservative Q-Learning} (SA-CQL), by building SA-PP upon representative CQL algorithm with the help of DualDICE for estimating discounted stationary state distribution ratios. Extensive experiments on standard offline RL benchmark show that SA-CQL outperforms the popular baselines on a large portion of benchmarks and attains the highest average return.

Taicai Chen, Yue Duan, Dong Li et al.

Variational Autoencoder based Bayesian Optimization (VAE-BO) has demonstrated its excellent performance in addressing high-dimensional structured optimization problems. However, current mainstream methods overlook the potential of utilizing a pool of unlabeled data to construct the latent space, while only concentrating on designing sophisticated models to leverage the labeled data. Despite their effective usage of labeled data, these methods often require extra network structures, additional procedure, resulting in computational inefficiency. To address this issue, we propose a novel method to effectively utilize unlabeled data with the guidance of labeled data. Specifically, we tailor the pseudo-labeling technique from semi-supervised learning to explicitly reveal the relative magnitudes of optimization objective values hidden within the unlabeled data. Based on this technique, we assign appropriate training weights to unlabeled data to enhance the construction of a discriminative latent space. Furthermore, we treat the VAE encoder and the Gaussian Process (GP) in Bayesian optimization as a unified deep kernel learning process, allowing the direct utilization of labeled data, which we term as Gaussian Process guidance. This directly and effectively integrates the goal of improving GP accuracy into the VAE training, thereby guiding the construction of the latent space. The extensive experiments demonstrate that our proposed method outperforms existing VAE-BO algorithms in various optimization scenarios. Our code will be published at https://github.com/TaicaiChen/PG-LBO.

Pancheng Wang, Shasha Li, Dong Li et al.

Automatically condensing multiple topic-related scientific papers into a succinct and concise summary is referred to as Multi-Document Scientific Summarization (MDSS). Currently, while commonly used abstractive MDSS methods can generate flexible and coherent summaries, the difficulty in handling global information and the lack of guidance during decoding still make it challenging to generate better summaries. To alleviate these two shortcomings, this paper introduces summary candidates into MDSS, utilizing the global information of the document set and additional guidance from the summary candidates to guide the decoding process. Our insights are twofold: Firstly, summary candidates can provide instructive information from both positive and negative perspectives, and secondly, selecting higher-quality candidates from multiple options contributes to producing better summaries. Drawing on the insights, we propose a summary candidates fusion framework -- Disentangling Instructive information from Ranked candidates (DIR) for MDSS. Specifically, DIR first uses a specialized pairwise comparison method towards multiple candidates to pick out those of higher quality. Then DIR disentangles the instructive information of summary candidates into positive and negative latent variables with Conditional Variational Autoencoder. These variables are further incorporated into the decoder to guide generation. We evaluate our approach with three different types of Transformer-based models and three different types of candidates, and consistently observe noticeable performance improvements according to automatic and human evaluation. More analyses further demonstrate the effectiveness of our model in handling global information and enhancing decoding controllability.

Hangyu Mao, Chao Wang, Xiaotian Hao et al.

The MineRL competition is designed for the development of reinforcement learning and imitation learning algorithms that can efficiently leverage human demonstrations to drastically reduce the number of environment interactions needed to solve the complex \emph{ObtainDiamond} task with sparse rewards. To address the challenge, in this paper, we present \textbf{SEIHAI}, a \textbf{S}ample-\textbf{e}ff\textbf{i}cient \textbf{H}ierarchical \textbf{AI}, that fully takes advantage of the human demonstrations and the task structure. Specifically, we split the task into several sequentially dependent subtasks, and train a suitable agent for each subtask using reinforcement learning and imitation learning. We further design a scheduler to select different agents for different subtasks automatically. SEIHAI takes the first place in the preliminary and final of the NeurIPS-2020 MineRL competition.

Haichao Zhu, Li Dong, Furu Wei et al.

The limited size of existing query-focused summarization datasets renders training data-driven summarization models challenging. Meanwhile, the manual construction of a query-focused summarization corpus is costly and time-consuming. In this paper, we use Wikipedia to automatically collect a large query-focused summarization dataset (named WIKIREF) of more than 280, 000 examples, which can serve as a means of data augmentation. We also develop a BERT-based query-focused summarization model (Q-BERT) to extract sentences from the documents as summaries. To better adapt a huge model containing millions of parameters to tiny benchmarks, we identify and fine-tune only a sparse subnetwork, which corresponds to a small fraction of the whole model parameters. Experimental results on three DUC benchmarks show that the model pre-trained on WIKIREF has already achieved reasonable performance. After fine-tuning on the specific benchmark datasets, the model with data augmentation outperforms strong comparison systems. Moreover, both our proposed Q-BERT model and subnetwork fine-tuning further improve the model performance. The dataset is publicly available at https://aka.ms/wikiref.