Mu Li

Zhuosheng Zhang, Aston Zhang, Mu Li et al.

Large language models (LLMs) have shown impressive performance on complex reasoning by leveraging chain-of-thought (CoT) prompting to generate intermediate reasoning chains as the rationale to infer the answer. However, existing CoT studies have primarily focused on the language modality. We propose Multimodal-CoT that incorporates language (text) and vision (images) modalities into a two-stage framework that separates rationale generation and answer inference. In this way, answer inference can leverage better generated rationales that are based on multimodal information. Experimental results on ScienceQA and A-OKVQA benchmark datasets show the effectiveness of our proposed approach. With Multimodal-CoT, our model under 1 billion parameters achieves state-of-the-art performance on the ScienceQA benchmark. Our analysis indicates that Multimodal-CoT offers the advantages of mitigating hallucination and enhancing convergence speed. Code is publicly available at https://github.com/amazon-science/mm-cot.

Zhuosheng Zhang, Aston Zhang, Mu Li et al.

Large language models (LLMs) can perform complex reasoning by generating intermediate reasoning steps. Providing these steps for prompting demonstrations is called chain-of-thought (CoT) prompting. CoT prompting has two major paradigms. One leverages a simple prompt like "Let's think step by step" to facilitate step-by-step thinking before answering a question. The other uses a few manual demonstrations one by one, each composed of a question and a reasoning chain that leads to an answer. The superior performance of the second paradigm hinges on the hand-crafting of task-specific demonstrations one by one. We show that such manual efforts may be eliminated by leveraging LLMs with the "Let's think step by step" prompt to generate reasoning chains for demonstrations one by one, i.e., let's think not just step by step, but also one by one. However, these generated chains often come with mistakes. To mitigate the effect of such mistakes, we find that diversity matters for automatically constructing demonstrations. We propose an automatic CoT prompting method: Auto-CoT. It samples questions with diversity and generates reasoning chains to construct demonstrations. On ten public benchmark reasoning tasks with GPT-3, Auto-CoT consistently matches or exceeds the performance of the CoT paradigm that requires manual designs of demonstrations. Code is available at https://github.com/amazon-research/auto-cot

Shuhuai Ren, Aston Zhang, Yi Zhu et al. · amazon-science, pku

This work proposes POMP, a prompt pre-training method for vision-language models. Being memory and computation efficient, POMP enables the learned prompt to condense semantic information for a rich set of visual concepts with over twenty-thousand classes. Once pre-trained, the prompt with a strong transferable ability can be directly plugged into a variety of visual recognition tasks including image classification, semantic segmentation, and object detection, to boost recognition performances in a zero-shot manner. Empirical evaluation shows that POMP achieves state-of-the-art performances on 21 datasets, e.g., 67.0% average accuracy on 10 classification datasets (+3.1% compared to CoOp) and 84.4 hIoU on open-vocabulary Pascal VOC segmentation (+6.9 compared to ZSSeg). Our code is available at https://github.com/amazon-science/prompt-pretraining.

Haotao Wang, Aston Zhang, Yi Zhu et al.

Existing out-of-distribution (OOD) detection methods are typically benchmarked on training sets with balanced class distributions. However, in real-world applications, it is common for the training sets to have long-tailed distributions. In this work, we first demonstrate that existing OOD detection methods commonly suffer from significant performance degradation when the training set is long-tail distributed. Through analysis, we posit that this is because the models struggle to distinguish the minority tail-class in-distribution samples, from the true OOD samples, making the tail classes more prone to be falsely detected as OOD. To solve this problem, we propose Partial and Asymmetric Supervised Contrastive Learning (PASCL), which explicitly encourages the model to distinguish between tail-class in-distribution samples and OOD samples. To further boost in-distribution classification accuracy, we propose Auxiliary Branch Finetuning, which uses two separate branches of BN and classification layers for anomaly detection and in-distribution classification, respectively. The intuition is that in-distribution and OOD anomaly data have different underlying distributions. Our method outperforms previous state-of-the-art method by $1.29\%$, $1.45\%$, $0.69\%$ anomaly detection false positive rate (FPR) and $3.24\%$, $4.06\%$, $7.89\%$ in-distribution classification accuracy on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT, respectively. Code and pre-trained models are available at https://github.com/amazon-research/long-tailed-ood-detection.

Yunhe Gao, Xingjian Shi, Yi Zhu et al. · amazon-science

Models should be able to adapt to unseen data during test-time to avoid performance drops caused by inevitable distribution shifts in real-world deployment scenarios. In this work, we tackle the practical yet challenging test-time adaptation (TTA) problem, where a model adapts to the target domain without accessing the source data. We propose a simple recipe called \textit{Data-efficient Prompt Tuning} (DePT) with two key ingredients. First, DePT plugs visual prompts into the vision Transformer and only tunes these source-initialized prompts during adaptation. We find such parameter-efficient finetuning can efficiently adapt the model representation to the target domain without overfitting to the noise in the learning objective. Second, DePT bootstraps the source representation to the target domain by memory bank-based online pseudo-labeling. A hierarchical self-supervised regularization specially designed for prompts is jointly optimized to alleviate error accumulation during self-training. With much fewer tunable parameters, DePT demonstrates not only state-of-the-art performance on major adaptation benchmarks VisDA-C, ImageNet-C, and DomainNet-126, but also superior data efficiency, i.e., adaptation with only 1\% or 10\% data without much performance degradation compared to 100\% data. In addition, DePT is also versatile to be extended to online or multi-source TTA settings.

Jiaao Chen, Aston Zhang, Xingjian Shi et al. · gatech

Parameter-efficient fine-tuning aims to achieve performance comparable to fine-tuning, using fewer trainable parameters. Several strategies (e.g., Adapters, prefix tuning, BitFit, and LoRA) have been proposed. However, their designs are hand-crafted separately, and it remains unclear whether certain design patterns exist for parameter-efficient fine-tuning. Thus, we present a parameter-efficient fine-tuning design paradigm and discover design patterns that are applicable to different experimental settings. Instead of focusing on designing another individual tuning strategy, we introduce parameter-efficient fine-tuning design spaces that parameterize tuning structures and tuning strategies. Specifically, any design space is characterized by four components: layer grouping, trainable parameter allocation, tunable groups, and strategy assignment. Starting from an initial design space, we progressively refine the space based on the model quality of each design choice and make greedy selection at each stage over these four components. We discover the following design patterns: (i) group layers in a spindle pattern; (ii) allocate the number of trainable parameters to layers uniformly; (iii) tune all the groups; (iv) assign proper tuning strategies to different groups. These design patterns result in new parameter-efficient fine-tuning methods. We show experimentally that these methods consistently and significantly outperform investigated parameter-efficient fine-tuning strategies across different backbone models and different tasks in natural language processing.

Jiaao Chen, Aston Zhang, Mu Li et al. · gatech

Diffusion models that are based on iterative denoising have been recently proposed and leveraged in various generation tasks like image generation. Whereas, as a way inherently built for continuous data, existing diffusion models still have some limitations in modeling discrete data, e.g., languages. For example, the generally used Gaussian noise can not handle the discrete corruption well, and the objectives in continuous spaces fail to be stable for textual data in the diffusion process especially when the dimension is high. To alleviate these issues, we introduce a novel diffusion model for language modeling, Masked-Diffuse LM, with lower training cost and better performances, inspired by linguistic features in languages. Specifically, we design a linguistic-informed forward process which adds corruptions to the text through strategically soft-masking to better noise the textual data. Also, we directly predict the categorical distribution with cross-entropy loss function in every diffusion step to connect the continuous space and discrete space in a more efficient and straightforward way. Through experiments on 5 controlled generation tasks, we demonstrate that our Masked-Diffuse LM can achieve better generation quality than the state-of-the-art diffusion models with better efficiency.

Likun Cai, Zhi Zhang, Yi Zhu et al.

Multiple datasets and open challenges for object detection have been introduced in recent years. To build more general and powerful object detection systems, in this paper, we construct a new large-scale benchmark termed BigDetection. Our goal is to simply leverage the training data from existing datasets (LVIS, OpenImages and Object365) with carefully designed principles, and curate a larger dataset for improved detector pre-training. Specifically, we generate a new taxonomy which unifies the heterogeneous label spaces from different sources. Our BigDetection dataset has 600 object categories and contains over 3.4M training images with 36M bounding boxes. It is much larger in multiple dimensions than previous benchmarks, which offers both opportunities and challenges. Extensive experiments demonstrate its validity as a new benchmark for evaluating different object detection methods, and its effectiveness as a pre-training dataset.

Xinnian Liang, Jing Li, Shuangzhi Wu et al.

Unsupervised summarization methods have achieved remarkable results by incorporating representations from pre-trained language models. However, existing methods fail to consider efficiency and effectiveness at the same time when the input document is extremely long. To tackle this problem, in this paper, we proposed an efficient Coarse-to-Fine Facet-Aware Ranking (C2F-FAR) framework for unsupervised long document summarization, which is based on the semantic block. The semantic block refers to continuous sentences in the document that describe the same facet. Specifically, we address this problem by converting the one-step ranking method into the hierarchical multi-granularity two-stage ranking. In the coarse-level stage, we propose a new segment algorithm to split the document into facet-aware semantic blocks and then filter insignificant blocks. In the fine-level stage, we select salient sentences in each block and then extract the final summary from selected sentences. We evaluate our framework on four long document summarization datasets: Gov-Report, BillSum, arXiv, and PubMed. Our C2F-FAR can achieve new state-of-the-art unsupervised summarization results on Gov-Report and BillSum. In addition, our method speeds up 4-28 times more than previous methods.\footnote{\url{https://github.com/xnliang98/c2f-far}}

Jiali Zeng, Yufan Jiang, Shuangzhi Wu et al.

Pretrained language models (PLMs) trained on large-scale unlabeled corpus are typically fine-tuned on task-specific downstream datasets, which have produced state-of-the-art results on various NLP tasks. However, the data discrepancy issue in domain and scale makes fine-tuning fail to efficiently capture task-specific patterns, especially in the low data regime. To address this issue, we propose Task-guided Disentangled Tuning (TDT) for PLMs, which enhances the generalization of representations by disentangling task-relevant signals from the entangled representations. For a given task, we introduce a learnable confidence model to detect indicative guidance from context, and further propose a disentangled regularization to mitigate the over-reliance problem. Experimental results on GLUE and CLUE benchmarks show that TDT gives consistently better results than fine-tuning with different PLMs, and extensive analysis demonstrates the effectiveness and robustness of our method. Code is available at https://github.com/lemon0830/TDT.

M Saiful Bari, Aston Zhang, Shuai Zheng et al.

Pre-trained large language models can efficiently interpolate human-written prompts in a natural way. Multitask prompted learning can help generalization through a diverse set of tasks at once, thus enhancing the potential for more effective downstream fine-tuning. To perform efficient multitask-inference in the same batch, parameter-efficient fine-tuning methods such as prompt tuning have been proposed. However, the existing prompt tuning methods may lack generalization. We propose SPT, a semi-parametric prompt tuning method for multitask prompted learning. The novel component of SPT is a memory bank from where memory prompts are retrieved based on discrete prompts. Extensive experiments, such as (i) fine-tuning a full language model with SPT on 31 different tasks from 8 different domains and evaluating zero-shot generalization on 9 heldout datasets under 5 NLP task categories and (ii) pretraining SPT on the GLUE datasets and evaluating fine-tuning on the SuperGLUE datasets, demonstrate effectiveness of SPT.

Yu Lu, Jiali Zeng, Jiajun Zhang et al.

Confidence estimation aims to quantify the confidence of the model prediction, providing an expectation of success. A well-calibrated confidence estimate enables accurate failure prediction and proper risk measurement when given noisy samples and out-of-distribution data in real-world settings. However, this task remains a severe challenge for neural machine translation (NMT), where probabilities from softmax distribution fail to describe when the model is probably mistaken. To address this problem, we propose an unsupervised confidence estimate learning jointly with the training of the NMT model. We explain confidence as how many hints the NMT model needs to make a correct prediction, and more hints indicate low confidence. Specifically, the NMT model is given the option to ask for hints to improve translation accuracy at the cost of some slight penalty. Then, we approximate their level of confidence by counting the number of hints the model uses. We demonstrate that our learned confidence estimate achieves high accuracy on extensive sentence/word-level quality estimation tasks. Analytical results verify that our confidence estimate can correctly assess underlying risk in two real-world scenarios: (1) discovering noisy samples and (2) detecting out-of-domain data. We further propose a novel confidence-based instance-specific label smoothing approach based on our learned confidence estimate, which outperforms standard label smoothing.

Cody Hao Yu, Haozheng Fan, Guangtai Huang et al.

As deep learning is pervasive in modern applications, many deep learning frameworks are presented for deep learning practitioners to develop and train DNN models rapidly. Meanwhile, as training large deep learning models becomes a trend in recent years, the training throughput and memory footprint are getting crucial. Accordingly, optimizing training workloads with compiler optimizations is inevitable and getting more and more attentions. However, existing deep learning compilers (DLCs) mainly target inference and do not incorporate holistic optimizations, such as automatic differentiation and automatic mixed precision, in training workloads. In this paper, we present RAF, a deep learning compiler for training. Unlike existing DLCs, RAF accepts a forward model and in-house generates a training graph. Accordingly, RAF is able to systematically consolidate graph optimizations for performance, memory and distributed training. In addition, to catch up to the state-of-the-art performance with hand-crafted kernel libraries as well as tensor compilers, RAF proposes an operator dialect mechanism to seamlessly integrate all possible kernel implementations. We demonstrate that by in-house training graph generation and operator dialect mechanism, we are able to perform holistic optimizations and achieve either better training throughput or larger batch size against PyTorch (eager and torchscript mode), XLA, and DeepSpeed for popular transformer models on GPUs.

Ruskin Raj Manku, Yuzhi Tang, Xingjian Shi et al.

Text-to-Speech (TTS) benchmarks often fail to capture how well models handle nuanced and semantically complex text. Building on $\textit{EmergentTTS}$, we introduce $\textit{EmergentTTS-Eval}$, a comprehensive benchmark covering six challenging TTS scenarios: emotions, paralinguistics, foreign words, syntactic complexity, complex pronunciation (e.g. URLs, formulas), and questions. Crucially, our framework automates both test-case generation and evaluation, making the benchmark easily extensible. Starting from a small set of human-written seed prompts, we iteratively extend them using LLMs to target specific structural, phonetic and prosodic challenges, resulting in 1,645 diverse test cases. Moreover, we employ a model-as-a-judge approach, using a Large Audio Language Model (LALM) to assess the speech across multiple dimensions such as expressed emotion, prosodic, intonational, and pronunciation accuracy. We evaluate state-of-the-art open-source and proprietary TTS systems, such as 11Labs, Deepgram, and OpenAI's 4o-mini-TTS, on EmergentTTS-Eval, demonstrating its ability to reveal fine-grained performance differences. Results show that the model-as-a-judge approach offers robust TTS assessment and a high correlation with human preferences. We open source the evaluation $\href{https://github.com/boson-ai/EmergentTTS-Eval-public}{code}$ and the $\href{https://huggingface.co/datasets/bosonai/EmergentTTS-Eval}{dataset}$.

Zhiqiang Tang, Yunhe Gao, Yi Zhu et al.

Traditional normalization techniques (e.g., Batch Normalization and Instance Normalization) generally and simplistically assume that training and test data follow the same distribution. As distribution shifts are inevitable in real-world applications, well-trained models with previous normalization methods can perform badly in new environments. Can we develop new normalization methods to improve generalization robustness under distribution shifts? In this paper, we answer the question by proposing CrossNorm and SelfNorm. CrossNorm exchanges channel-wise mean and variance between feature maps to enlarge training distribution, while SelfNorm uses attention to recalibrate the statistics to bridge gaps between training and test distributions. CrossNorm and SelfNorm can complement each other, though exploring different directions in statistics usage. Extensive experiments on different fields (vision and language), tasks (classification and segmentation), settings (supervised and semi-supervised), and distribution shift types (synthetic and natural) show the effectiveness. Code is available at https://github.com/amazon-research/crossnorm-selfnorm

Jingbo Lu, Leheng Zhang, Xingyu Zhou et al.

Learned image compression methods have attracted great research interest and exhibited superior rate-distortion performance to the best classical image compression standards of the present. The entropy model plays a key role in learned image compression, which estimates the probability distribution of the latent representation for further entropy coding. Most existing methods employed hyper-prior and auto-regressive architectures to form their entropy models. However, they only aimed to explore the internal dependencies of latent representation while neglecting the importance of extracting prior from training data. In this work, we propose a novel entropy model named Dictionary-based Cross Attention Entropy model, which introduces a learnable dictionary to summarize the typical structures occurring in the training dataset to enhance the entropy model. Extensive experimental results have demonstrated that the proposed model strikes a better balance between performance and latency, achieving state-of-the-art results on various benchmark datasets.

Pengfei Yu, Dongming Shen, Silin Meng et al.

We present RPGBench, the first benchmark designed to evaluate large language models (LLMs) as text-based role-playing game (RPG) engines. RPGBench comprises two core tasks: Game Creation (GC) and Game Simulation (GS). In GC, an LLM must craft a valid and playable RPG world using a structured event-state representation, ensuring logical coherence and proper termination conditions. In GS, the LLM simulates interactive gameplay across multiple rounds while consistently updating states and enforcing game rules. To comprehensively assess performance, RPGBench integrates objective and subjective evaluation methodologies. Objective measures verify adherence to event mechanics and check variable updates without requiring human intervention. Subjective measures, such as content interestingness, action quality, and role-playing capability, are evaluated via an LLM-as-a-judge framework, where a strong LLM grades each candidate's outputs. Empirical results demonstrate that state-of-the-art LLMs can produce engaging stories but often struggle to implement consistent, verifiable game mechanics, particularly in long or complex scenarios. By combining structured, rule-based assessments with LLM-based judgments, RPGBench provides a new standard for evaluating how well LLMs can balance creativity, coherence, and complexity in text-based RPGs, opening avenues for more immersive and controllable interactive storytelling.

Qingao Dong, Mengfei Wang, Hengzhi Zhang et al.

Large language model (LLM) agents have recently shown strong performance on repository-level issue resolution, but existing systems are almost exclusively designed for Python and rely heavily on lexical retrieval and shallow code navigation. These approaches transfer poorly to C++ projects, where overloaded identifiers, nested namespaces, template instantiations, and deep control-flow structures make context retrieval and fault localization substantially more difficult. As a result, state-of-the-art Python-oriented agents show a drastic performance drop on the C++ subset of MultiSWE-bench. We introduce INFCODE-C++, the first C++-aware autonomous system for end-to-end issue resolution. The system combines two complementary retrieval mechanisms -- semantic code-intent retrieval and deterministic AST-structured querying -- to construct accurate, language-aware context for repair.These components enable precise localization and robust patch synthesis in large, statically typed C++ repositories. Evaluated on the \texttt{MultiSWE-bench-CPP} benchmark, INFCODE-C++ achieves a resolution rate of 25.58\%, outperforming the strongest prior agent by 10.85 percentage points and more than doubling the performance of MSWE-agent. Ablation and behavioral studies further demonstrate the critical role of semantic retrieval, structural analysis, and accurate reproduction in C++ issue resolution. INFCODE-C++ highlights the need for language-aware reasoning in multi-language software agents and establishes a foundation for future research on scalable, LLM-driven repair for complex, statically typed ecosystems.

Yin Wang, Mu li, Zhiying Leng et al.

We introduce MOST, a novel motion diffusion model via temporal clip Banzhaf interaction, aimed at addressing the persistent challenge of generating human motion from rare language prompts. While previous approaches struggle with coarse-grained matching and overlook important semantic cues due to motion redundancy, our key insight lies in leveraging fine-grained clip relationships to mitigate these issues. MOST's retrieval stage presents the first formulation of its kind - temporal clip Banzhaf interaction - which precisely quantifies textual-motion coherence at the clip level. This facilitates direct, fine-grained text-to-motion clip matching and eliminates prevalent redundancy. In the generation stage, a motion prompt module effectively utilizes retrieved motion clips to produce semantically consistent movements. Extensive evaluations confirm that MOST achieves state-of-the-art text-to-motion retrieval and generation performance by comprehensively addressing previous challenges, as demonstrated through quantitative and qualitative results highlighting its effectiveness, especially for rare prompts.

Yuxin Ren, Zihan Zhong, Xingjian Shi et al.

It has been commonly observed that a teacher model with superior performance does not necessarily result in a stronger student, highlighting a discrepancy between current teacher training practices and effective knowledge transfer. In order to enhance the guidance of the teacher training process, we introduce the concept of distillation influence to determine the impact of distillation from each training sample on the student's generalization ability. In this paper, we propose Learning Good Teacher Matters (LGTM), an efficient training technique for incorporating distillation influence into the teacher's learning process. By prioritizing samples that are likely to enhance the student's generalization ability, our LGTM outperforms 10 common knowledge distillation baselines on 6 text classification tasks in the GLUE benchmark.

Shengju Qian, Hao Shao, Yi Zhu et al.

The transformer architectures, based on self-attention mechanism and convolution-free design, recently found superior performance and booming applications in computer vision. However, the discontinuous patch-wise tokenization process implicitly introduces jagged artifacts into attention maps, arising the traditional problem of aliasing for vision transformers. Aliasing effect occurs when discrete patterns are used to produce high frequency or continuous information, resulting in the indistinguishable distortions. Recent researches have found that modern convolution networks still suffer from this phenomenon. In this work, we analyze the uncharted problem of aliasing in vision transformer and explore to incorporate anti-aliasing properties. Specifically, we propose a plug-and-play Aliasing-Reduction Module(ARM) to alleviate the aforementioned issue. We investigate the effectiveness and generalization of the proposed method across multiple tasks and various vision transformer families. This lightweight design consistently attains a clear boost over several famous structures. Furthermore, our module also improves data efficiency and robustness of vision transformers.

Haoyu He, Xingjian Shi, Jonas Mueller et al.

We aim to identify how different components in the KD pipeline affect the resulting performance and how much the optimal KD pipeline varies across different datasets/tasks, such as the data augmentation policy, the loss function, and the intermediate representation for transferring the knowledge between teacher and student. To tease apart their effects, we propose Distiller, a meta KD framework that systematically combines a broad range of techniques across different stages of the KD pipeline, which enables us to quantify each component's contribution. Within Distiller, we unify commonly used objectives for distillation of intermediate representations under a universal mutual information (MI) objective and propose a class of MI-$α$ objective functions with better bias/variance trade-off for estimating the MI between the teacher and the student. On a diverse set of NLP datasets, the best Distiller configurations are identified via large-scale hyperparameter optimization. Our experiments reveal the following: 1) the approach used to distill the intermediate representations is the most important factor in KD performance, 2) among different objectives for intermediate distillation, MI-$α$ performs the best, and 3) data augmentation provides a large boost for small training datasets or small student networks. Moreover, we find that different datasets/tasks prefer different KD algorithms, and thus propose a simple AutoDistiller algorithm that can recommend a good KD pipeline for a new dataset.

Fangrui Zhu, Yi Zhu, Li Zhang et al.

Semantic segmentation is a challenging problem due to difficulties in modeling context in complex scenes and class confusions along boundaries. Most literature either focuses on context modeling or boundary refinement, which is less generalizable in open-world scenarios. In this work, we advocate a unified framework(UN-EPT) to segment objects by considering both context information and boundary artifacts. We first adapt a sparse sampling strategy to incorporate the transformer-based attention mechanism for efficient context modeling. In addition, a separate spatial branch is introduced to capture image details for boundary refinement. The whole model can be trained in an end-to-end manner. We demonstrate promising performance on three popular benchmarks for semantic segmentation with low memory footprint. Code will be released soon.

Yufan Jiang, Shuangzhi Wu, Jing Gong et al.

Multi-choice Machine Reading Comprehension (MMRC) aims to select the correct answer from a set of options based on a given passage and question. Due to task specific of MMRC, it is non-trivial to transfer knowledge from other MRC tasks such as SQuAD, Dream. In this paper, we simply reconstruct multi-choice to single-choice by training a binary classification to distinguish whether a certain answer is correct. Then select the option with the highest confidence score. We construct our model upon ALBERT-xxlarge model and estimate it on the RACE dataset. During training, We adopt AutoML strategy to tune better parameters. Experimental results show that the single-choice is better than multi-choice. In addition, by transferring knowledge from other kinds of MRC tasks, our model achieves a new state-of-the-art results in both single and ensemble settings.

Yuwei Hu, Zihao Ye, Minjie Wang et al.

Graph neural networks (GNNs) are gaining increasing popularity as a promising approach to machine learning on graphs. Unlike traditional graph workloads where each vertex/edge is associated with a scalar, GNNs attach a feature tensor to each vertex/edge. This additional feature dimension, along with consequently more complex vertex- and edge-wise computations, has enormous implications on locality and parallelism, which existing graph processing systems fail to exploit. This paper proposes FeatGraph to accelerate GNN workloads by co-optimizing graph traversal and feature dimension computation. FeatGraph provides a flexible programming interface to express diverse GNN models by composing coarse-grained sparse templates with fine-grained user-defined functions (UDFs) on each vertex/edge. FeatGraph incorporates optimizations for graph traversal into the sparse templates and allows users to specify optimizations for UDFs with a feature dimension schedule (FDS). FeatGraph speeds up end-to-end GNN training and inference by up to 32x on CPU and 7x on GPU.

Haichen Shen, Jared Roesch, Zhi Chen et al.

Modern deep neural networks increasingly make use of features such as dynamic control flow, data structures and dynamic tensor shapes. Existing deep learning systems focus on optimizing and executing static neural networks which assume a pre-determined model architecture and input data shapes--assumptions which are violated by dynamic neural networks. Therefore, executing dynamic models with deep learning systems is currently both inflexible and sub-optimal, if not impossible. Optimizing dynamic neural networks is more challenging than static neural networks; optimizations must consider all possible execution paths and tensor shapes. This paper proposes Nimble, a high-performance and flexible system to optimize, compile, and execute dynamic neural networks on multiple platforms. Nimble handles model dynamism by introducing a dynamic type system, a set of dynamism-oriented optimizations, and a light-weight virtual machine runtime. Our evaluation demonstrates that Nimble outperforms state-of-the-art deep learning frameworks and runtime systems for dynamic neural networks by up to 20x on hardware platforms including Intel CPUs, ARM CPUs, and Nvidia GPUs.

Yi Zhu, Zhongyue Zhang, Chongruo Wu et al.

Deep learning usually achieves the best results with complete supervision. In the case of semantic segmentation, this means that large amounts of pixelwise annotations are required to learn accurate models. In this paper, we show that we can obtain state-of-the-art results using a semi-supervised approach, specifically a self-training paradigm. We first train a teacher model on labeled data, and then generate pseudo labels on a large set of unlabeled data. Our robust training framework can digest human-annotated and pseudo labels jointly and achieve top performances on Cityscapes, CamVid and KITTI datasets while requiring significantly less supervision. We also demonstrate the effectiveness of self-training on a challenging cross-domain generalization task, outperforming conventional finetuning method by a large margin. Lastly, to alleviate the computational burden caused by the large amount of pseudo labels, we propose a fast training schedule to accelerate the training of segmentation models by up to 2x without performance degradation.

Hang Zhang, Chongruo Wu, Zhongyue Zhang et al.

It is well known that featuremap attention and multi-path representation are important for visual recognition. In this paper, we present a modularized architecture, which applies the channel-wise attention on different network branches to leverage their success in capturing cross-feature interactions and learning diverse representations. Our design results in a simple and unified computation block, which can be parameterized using only a few variables. Our model, named ResNeSt, outperforms EfficientNet in accuracy and latency trade-off on image classification. In addition, ResNeSt has achieved superior transfer learning results on several public benchmarks serving as the backbone, and has been adopted by the winning entries of COCO-LVIS challenge. The source code for complete system and pretrained models are publicly available.

Jian Guo, He He, Tong He et al.

We present GluonCV and GluonNLP, the deep learning toolkits for computer vision and natural language processing based on Apache MXNet (incubating). These toolkits provide state-of-the-art pre-trained models, training scripts, and training logs, to facilitate rapid prototyping and promote reproducible research. We also provide modular APIs with flexible building blocks to enable efficient customization. Leveraging the MXNet ecosystem, the deep learning models in GluonCV and GluonNLP can be deployed onto a variety of platforms with different programming languages. The Apache 2.0 license has been adopted by GluonCV and GluonNLP to allow for software distribution, modification, and usage.

Haibin Lin, Hang Zhang, Yifei Ma et al.

With an increasing demand for training powers for deep learning algorithms and the rapid growth of computation resources in data centers, it is desirable to dynamically schedule different distributed deep learning tasks to maximize resource utilization and reduce cost. In this process, different tasks may receive varying numbers of machines at different time, a setting we call elastic distributed training. Despite the recent successes in large mini-batch distributed training, these methods are rarely tested in elastic distributed training environments and suffer degraded performance in our experiments, when we adjust the learning rate linearly immediately with respect to the batch size. One difficulty we observe is that the noise in the stochastic momentum estimation is accumulated over time and will have delayed effects when the batch size changes. We therefore propose to smoothly adjust the learning rate over time to alleviate the influence of the noisy momentum estimation. Our experiments on image classification, object detection and semantic segmentation have demonstrated that our proposed Dynamic SGD method achieves stabilized performance when varying the number of GPUs from 8 to 128. We also provide theoretical understanding on the optimality of linear learning rate scheduling and the effects of stochastic momentum.

Zhi Zhang, Tong He, Hang Zhang et al.

Training heuristics greatly improve various image classification model accuracies~\cite{he2018bag}. Object detection models, however, have more complex neural network structures and optimization targets. The training strategies and pipelines dramatically vary among different models. In this works, we explore training tweaks that apply to various models including Faster R-CNN and YOLOv3. These tweaks do not change the model architectures, therefore, the inference costs remain the same. Our empirical results demonstrate that, however, these freebies can improve up to 5% absolute precision compared to state-of-the-art baselines.

Tong He, Zhi Zhang, Hang Zhang et al.

Much of the recent progress made in image classification research can be credited to training procedure refinements, such as changes in data augmentations and optimization methods. In the literature, however, most refinements are either briefly mentioned as implementation details or only visible in source code. In this paper, we will examine a collection of such refinements and empirically evaluate their impact on the final model accuracy through ablation study. We will show that, by combining these refinements together, we are able to improve various CNN models significantly. For example, we raise ResNet-50's top-1 validation accuracy from 75.3% to 79.29% on ImageNet. We will also demonstrate that improvement on image classification accuracy leads to better transfer learning performance in other application domains such as object detection and semantic segmentation.

Zhirui Zhang, Shuo Ren, Shujie Liu et al.

Language style transferring rephrases text with specific stylistic attributes while preserving the original attribute-independent content. One main challenge in learning a style transfer system is a lack of parallel data where the source sentence is in one style and the target sentence in another style. With this constraint, in this paper, we adapt unsupervised machine translation methods for the task of automatic style transfer. We first take advantage of style-preference information and word embedding similarity to produce pseudo-parallel data with a statistical machine translation (SMT) framework. Then the iterative back-translation approach is employed to jointly train two neural machine translation (NMT) based transfer systems. To control the noise generated during joint training, a style classifier is introduced to guarantee the accuracy of style transfer and penalize bad candidates in the generated pseudo data. Experiments on benchmark datasets show that our proposed method outperforms previous state-of-the-art models in terms of both accuracy of style transfer and quality of input-output correspondence.

Zhirui Zhang, Shuangzhi Wu, Shujie Liu et al.

Although Neural Machine Translation (NMT) has achieved remarkable progress in the past several years, most NMT systems still suffer from a fundamental shortcoming as in other sequence generation tasks: errors made early in generation process are fed as inputs to the model and can be quickly amplified, harming subsequent sequence generation. To address this issue, we propose a novel model regularization method for NMT training, which aims to improve the agreement between translations generated by left-to-right (L2R) and right-to-left (R2L) NMT decoders. This goal is achieved by introducing two Kullback-Leibler divergence regularization terms into the NMT training objective to reduce the mismatch between output probabilities of L2R and R2L models. In addition, we also employ a joint training strategy to allow L2R and R2L models to improve each other in an interactive update process. Experimental results show that our proposed method significantly outperforms state-of-the-art baselines on Chinese-English and English-German translation tasks.

Shuo Ren, Wenhu Chen, Shujie Liu et al.

Neural Machine Translation (NMT) performs poor on the low-resource language pair $(X,Z)$, especially when $Z$ is a rare language. By introducing another rich language $Y$, we propose a novel triangular training architecture (TA-NMT) to leverage bilingual data $(Y,Z)$ (may be small) and $(X,Y)$ (can be rich) to improve the translation performance of low-resource pairs. In this triangular architecture, $Z$ is taken as the intermediate latent variable, and translation models of $Z$ are jointly optimized with a unified bidirectional EM algorithm under the goal of maximizing the translation likelihood of $(X,Y)$. Empirical results demonstrate that our method significantly improves the translation quality of rare languages on MultiUN and IWSLT2012 datasets, and achieves even better performance combining back-translation methods.

Hany Hassan, Anthony Aue, Chang Chen et al.

Machine translation has made rapid advances in recent years. Millions of people are using it today in online translation systems and mobile applications in order to communicate across language barriers. The question naturally arises whether such systems can approach or achieve parity with human translations. In this paper, we first address the problem of how to define and accurately measure human parity in translation. We then describe Microsoft's machine translation system and measure the quality of its translations on the widely used WMT 2017 news translation task from Chinese to English. We find that our latest neural machine translation system has reached a new state-of-the-art, and that the translation quality is at human parity when compared to professional human translations. We also find that it significantly exceeds the quality of crowd-sourced non-professional translations.

Zhirui Zhang, Shujie Liu, Mu Li et al.

Monolingual data have been demonstrated to be helpful in improving translation quality of both statistical machine translation (SMT) systems and neural machine translation (NMT) systems, especially in resource-poor or domain adaptation tasks where parallel data are not rich enough. In this paper, we propose a novel approach to better leveraging monolingual data for neural machine translation by jointly learning source-to-target and target-to-source NMT models for a language pair with a joint EM optimization method. The training process starts with two initial NMT models pre-trained on parallel data for each direction, and these two models are iteratively updated by incrementally decreasing translation losses on training data. In each iteration step, both NMT models are first used to translate monolingual data from one language to the other, forming pseudo-training data of the other NMT model. Then two new NMT models are learnt from parallel data together with the pseudo training data. Both NMT models are expected to be improved and better pseudo-training data can be generated in next step. Experiment results on Chinese-English and English-German translation tasks show that our approach can simultaneously improve translation quality of source-to-target and target-to-source models, significantly outperforming strong baseline systems which are enhanced with monolingual data for model training including back-translation.

Wenhu Chen, Guanlin Li, Shuo Ren et al.

In order to alleviate data sparsity and overfitting problems in maximum likelihood estimation (MLE) for sequence prediction tasks, we propose the Generative Bridging Network (GBN), in which a novel bridge module is introduced to assist the training of the sequence prediction model (the generator network). Unlike MLE directly maximizing the conditional likelihood, the bridge extends the point-wise ground truth to a bridge distribution conditioned on it, and the generator is optimized to minimize their KL-divergence. Three different GBNs, namely uniform GBN, language-model GBN and coaching GBN, are proposed to penalize confidence, enhance language smoothness and relieve learning burden. Experiments conducted on two recognized sequence prediction tasks (machine translation and abstractive text summarization) show that our proposed GBNs can yield significant improvements over strong baselines. Furthermore, by analyzing samples drawn from different bridges, expected influences on the generator are verified.

Bing Xu, Ruitong Huang, Mu Li

In this paper, we revise two commonly used saturated functions, the logistic sigmoid and the hyperbolic tangent (tanh). We point out that, besides the well-known non-zero centered property, slope of the activation function near the origin is another possible reason making training deep networks with the logistic function difficult to train. We demonstrate that, with proper rescaling, the logistic sigmoid achieves comparable results with tanh. Then following the same argument, we improve tahn by penalizing in the negative part. We show that "penalized tanh" is comparable and even outperforms the state-of-the-art non-saturated functions including ReLU and leaky ReLU on deep convolution neural networks. Our results contradict to the conclusion of previous works that the saturation property causes the slow convergence. It suggests further investigation is necessary to better understand activation functions in deep architectures.

Travis Dick, Mu Li, Venkata Krishna Pillutla et al.

In distributed machine learning, data is dispatched to multiple machines for processing. Motivated by the fact that similar data points often belong to the same or similar classes, and more generally, classification rules of high accuracy tend to be "locally simple but globally complex" (Vapnik & Bottou 1993), we propose data dependent dispatching that takes advantage of such structure. We present an in-depth analysis of this model, providing new algorithms with provable worst-case guarantees, analysis proving existing scalable heuristics perform well in natural non worst-case conditions, and techniques for extending a dispatching rule from a small sample to the entire distribution. We overcome novel technical challenges to satisfy important conditions for accurate distributed learning, including fault tolerance and balancedness. We empirically compare our approach with baselines based on random partitioning, balanced partition trees, and locality sensitive hashing, showing that we achieve significantly higher accuracy on both synthetic and real world image and advertising datasets. We also demonstrate that our technique strongly scales with the available computing power.

Aaron Q. Li, Amr Ahmed, Mu Li et al.

Latent variable models have accumulated a considerable amount of interest from the industry and academia for their versatility in a wide range of applications. A large amount of effort has been made to develop systems that is able to extend the systems to a large scale, in the hope to make use of them on industry scale data. In this paper, we describe a system that operates at a scale orders of magnitude higher than previous works, and an order of magnitude faster than state-of-the-art system at the same scale, at the same time showing more robustness and more accurate results. Our system uses a number of advances in distributed inference: high performance in synchronization of sufficient statistics with relaxed consistency model; fast sampling, using the Metropolis-Hastings-Walker method to overcome dense generative models; statistical modeling, moving beyond Latent Dirichlet Allocation (LDA) to Pitman-Yor distributions (PDP) and Hierarchical Dirichlet Process (HDP) models; sophisticated parameter projection schemes, to resolve the conflicts within the constraint between parameters arising from the relaxed consistency model. This work significantly extends the domain of applicability of what is commonly known as the Parameter Server. We obtain results with up to hundreds billion oftokens, thousands of topics, and a vocabulary of a few million token-types, using up to 60,000 processor cores operating on a production cluster of a large Internet company. This demonstrates the feasibility to scale to problems orders of magnitude larger than any previously published work.

Suvrit Sra, Adams Wei Yu, Mu Li et al.

We study distributed stochastic convex optimization under the delayed gradient model where the server nodes perform parameter updates, while the worker nodes compute stochastic gradients. We discuss, analyze, and experiment with a setup motivated by the behavior of real-world distributed computation networks, where the machines are differently slow at different time. Therefore, we allow the parameter updates to be sensitive to the actual delays experienced, rather than to worst-case bounds on the maximum delay. This sensitivity leads to larger stepsizes, that can help gain rapid initial convergence without having to wait too long for slower machines, while maintaining the same asymptotic complexity. We obtain encouraging improvements to overall convergence for distributed experiments on real datasets with up to billions of examples and features.

Mu Li, Dave G. Andersen, Alexander J. Smola

Distributed computing excels at processing large scale data, but the communication cost for synchronizing the shared parameters may slow down the overall performance. Fortunately, the interactions between parameter and data in many problems are sparse, which admits efficient partition in order to reduce the communication overhead. In this paper, we formulate data placement as a graph partitioning problem. We propose a distributed partitioning algorithm. We give both theoretical guarantees and a highly efficient implementation. We also provide a highly efficient implementation of the algorithm and demonstrate its promising results on both text datasets and social networks. We show that the proposed algorithm leads to 1.6x speedup of a state-of-the-start distributed machine learning system by eliminating 90\% of the network communication.

Bing Xu, Naiyan Wang, Tianqi Chen et al.

In this paper we investigate the performance of different types of rectified activation functions in convolutional neural network: standard rectified linear unit (ReLU), leaky rectified linear unit (Leaky ReLU), parametric rectified linear unit (PReLU) and a new randomized leaky rectified linear units (RReLU). We evaluate these activation function on standard image classification task. Our experiments suggest that incorporating a non-zero slope for negative part in rectified activation units could consistently improve the results. Thus our findings are negative on the common belief that sparsity is the key of good performance in ReLU. Moreover, on small scale dataset, using deterministic negative slope or learning it are both prone to overfitting. They are not as effective as using their randomized counterpart. By using RReLU, we achieved 75.68\% accuracy on CIFAR-100 test set without multiple test or ensemble.

Jiajun Zhang, Shujie Liu, Mu Li et al.

Language model is one of the most important modules in statistical machine translation and currently the word-based language model dominants this community. However, many translation models (e.g. phrase-based models) generate the target language sentences by rendering and compositing the phrases rather than the words. Thus, it is much more reasonable to model dependency between phrases, but few research work succeed in solving this problem. In this paper, we tackle this problem by designing a novel phrase-based language model which attempts to solve three key sub-problems: 1, how to define a phrase in language model; 2, how to determine the phrase boundary in the large-scale monolingual data in order to enlarge the training set; 3, how to alleviate the data sparsity problem due to the huge vocabulary size of phrases. By carefully handling these issues, the extensive experiments on Chinese-to-English translation show that our phrase-based language model can significantly improve the translation quality by up to +1.47 absolute BLEU score.