Jia Li

Dingfeng Shi, Yujie Zhong, Qiong Cao et al.

This work aims at advancing temporal action detection (TAD) using an encoder-decoder framework with action queries, similar to DETR, which has shown great success in object detection. However, the framework suffers from several problems if directly applied to TAD: the insufficient exploration of inter-query relation in the decoder, the inadequate classification training due to a limited number of training samples, and the unreliable classification scores at inference. To this end, we first propose a relational attention mechanism in the decoder, which guides the attention among queries based on their relations. Moreover, we propose two losses to facilitate and stabilize the training of action classification. Lastly, we propose to predict the localization quality of each action query at inference in order to distinguish high-quality queries. The proposed method, named ReAct, achieves the state-of-the-art performance on THUMOS14, with much lower computational costs than previous methods. Besides, extensive ablation studies are conducted to verify the effectiveness of each proposed component. The code is available at https://github.com/sssste/React.

Peng Dai, Xin Yu, Lan Ma et al.

Moire patterns, appearing as color distortions, severely degrade image and video qualities when filming a screen with digital cameras. Considering the increasing demands for capturing videos, we study how to remove such undesirable moire patterns in videos, namely video demoireing. To this end, we introduce the first hand-held video demoireing dataset with a dedicated data collection pipeline to ensure spatial and temporal alignments of captured data. Further, a baseline video demoireing model with implicit feature space alignment and selective feature aggregation is developed to leverage complementary information from nearby frames to improve frame-level video demoireing. More importantly, we propose a relation-based temporal consistency loss to encourage the model to learn temporal consistency priors directly from ground-truth reference videos, which facilitates producing temporally consistent predictions and effectively maintains frame-level qualities. Extensive experiments manifest the superiority of our model. Code is available at \url{https://daipengwa.github.io/VDmoire_ProjectPage/}.

Lanqing Li, Liang Zeng, Ziqi Gao et al.

The last decade has witnessed a prosperous development of computational methods and dataset curation for AI-aided drug discovery (AIDD). However, real-world pharmaceutical datasets often exhibit highly imbalanced distribution, which is overlooked by the current literature but may severely compromise the fairness and generalization of machine learning applications. Motivated by this observation, we introduce ImDrug, a comprehensive benchmark with an open-source Python library which consists of 4 imbalance settings, 11 AI-ready datasets, 54 learning tasks and 16 baseline algorithms tailored for imbalanced learning. It provides an accessible and customizable testbed for problems and solutions spanning a broad spectrum of the drug discovery pipeline such as molecular modeling, drug-target interaction and retrosynthesis. We conduct extensive empirical studies with novel evaluation metrics, to demonstrate that the existing algorithms fall short of solving medicinal and pharmaceutical challenges in the data imbalance scenario. We believe that ImDrug opens up avenues for future research and development, on real-world challenges at the intersection of AIDD and deep imbalanced learning.

Dingfeng Shi, Yujie Zhong, Qiong Cao et al.

In this paper, we present a one-stage framework TriDet for temporal action detection. Existing methods often suffer from imprecise boundary predictions due to the ambiguous action boundaries in videos. To alleviate this problem, we propose a novel Trident-head to model the action boundary via an estimated relative probability distribution around the boundary. In the feature pyramid of TriDet, we propose an efficient Scalable-Granularity Perception (SGP) layer to mitigate the rank loss problem of self-attention that takes place in the video features and aggregate information across different temporal granularities. Benefiting from the Trident-head and the SGP-based feature pyramid, TriDet achieves state-of-the-art performance on three challenging benchmarks: THUMOS14, HACS and EPIC-KITCHEN 100, with lower computational costs, compared to previous methods. For example, TriDet hits an average mAP of $69.3\%$ on THUMOS14, outperforming the previous best by $2.5\%$, but with only $74.6\%$ of its latency. The code is released to https://github.com/sssste/TriDet.

Haojie Zhang, Ge Li, Jia Li et al.

Large-scale pre-trained language models have achieved impressive results on a wide range of downstream tasks recently. However, fine-tuning an extremely large-scale pre-trained language model on limited target datasets is often plagued by overfitting and representation degradation. In this paper, we propose a Dynamic Parameter Selection (DPS) algorithm for the large-scale pre-trained models during fine-tuning, which adaptively selects a more promising subnetwork to perform staging updates based on gradients of back-propagation. Experiments on the GLUE benchmark show that DPS outperforms previous fine-tuning methods in terms of overall performance and stability, and consistently achieves better results with variable pre-trained language models. In addition, DPS brings a large magnitude of improvement in out-of-domain transferring experiments and low-resource scenarios, which shows that it can maintain stable general contextual features and reduce the representation collapse. We release our code at https://github.com/ZhangHaojie077/DPS

Wenzhuo Yang, Jia Li, Caiming Xiong et al. · salesforce

Counterfactual explanation is an important Explainable AI technique to explain machine learning predictions. Despite being studied actively, existing optimization-based methods often assume that the underlying machine-learning model is differentiable and treat categorical attributes as continuous ones, which restricts their real-world applications when categorical attributes have many different values or the model is non-differentiable. To make counterfactual explanation suitable for real-world applications, we propose a novel framework of Model-Agnostic Counterfactual Explanation (MACE), which adopts a newly designed pipeline that can efficiently handle non-differentiable machine-learning models on a large number of feature values. in our MACE approach, we propose a novel RL-based method for finding good counterfactual examples and a gradient-less descent method for improving proximity. Experiments on public datasets validate the effectiveness with better validity, sparsity and proximity.

Xin Yu, Peng Dai, Wenbo Li et al.

With the rapid development of mobile devices, modern widely-used mobile phones typically allow users to capture 4K resolution (i.e., ultra-high-definition) images. However, for image demoireing, a challenging task in low-level vision, existing works are generally carried out on low-resolution or synthetic images. Hence, the effectiveness of these methods on 4K resolution images is still unknown. In this paper, we explore moire pattern removal for ultra-high-definition images. To this end, we propose the first ultra-high-definition demoireing dataset (UHDM), which contains 5,000 real-world 4K resolution image pairs, and conduct a benchmark study on current state-of-the-art methods. Further, we present an efficient baseline model ESDNet for tackling 4K moire images, wherein we build a semantic-aligned scale-aware module to address the scale variation of moire patterns. Extensive experiments manifest the effectiveness of our approach, which outperforms state-of-the-art methods by a large margin while being much more lightweight. Code and dataset are available at https://xinyu-andy.github.io/uhdm-page.

Yang Liu, Jiashun Cheng, Haihong Zhao et al.

Graph Neural Networks (GNNs) with equivariant properties have emerged as powerful tools for modeling complex dynamics of multi-object physical systems. However, their generalization ability is limited by the inadequate consideration of physical inductive biases: (1) Existing studies overlook the continuity of transitions among system states, opting to employ several discrete transformation layers to learn the direct mapping between two adjacent states; (2) Most models only account for first-order velocity information, despite the fact that many physical systems are governed by second-order motion laws. To incorporate these inductive biases, we propose the Second-order Equivariant Graph Neural Ordinary Differential Equation (SEGNO). Specifically, we show how the second-order continuity can be incorporated into GNNs while maintaining the equivariant property. Furthermore, we offer theoretical insights into SEGNO, highlighting that it can learn a unique trajectory between adjacent states, which is crucial for model generalization. Additionally, we prove that the discrepancy between this learned trajectory of SEGNO and the true trajectory is bounded. Extensive experiments on complex dynamical systems including molecular dynamics and motion capture demonstrate that our model yields a significant improvement over the state-of-the-art baselines.

Chenxi Xie, Changqun Xia, Mingcan Ma et al.

Recent salient object detection (SOD) methods based on deep neural network have achieved remarkable performance. However, most of existing SOD models designed for low-resolution input perform poorly on high-resolution images due to the contradiction between the sampling depth and the receptive field size. Aiming at resolving this contradiction, we propose a novel one-stage framework called Pyramid Grafting Network (PGNet), using transformer and CNN backbone to extract features from different resolution images independently and then graft the features from transformer branch to CNN branch. An attention-based Cross-Model Grafting Module (CMGM) is proposed to enable CNN branch to combine broken detailed information more holistically, guided by different source feature during decoding process. Moreover, we design an Attention Guided Loss (AGL) to explicitly supervise the attention matrix generated by CMGM to help the network better interact with the attention from different models. We contribute a new Ultra-High-Resolution Saliency Detection dataset UHRSD, containing 5,920 images at 4K-8K resolutions. To our knowledge, it is the largest dataset in both quantity and resolution for high-resolution SOD task, which can be used for training and testing in future research. Sufficient experiments on UHRSD and widely-used SOD datasets demonstrate that our method achieves superior performance compared to the state-of-the-art methods.

Jia Li, Ge Li, Chongyang Tao et al.

Large language models (LLMs) have shown impressive in-context learning (ICL) ability in code generation. LLMs take a prompt consisting of requirement-code examples and a new requirement as input, and output new programs. Existing studies have found that ICL is highly dominated by the examples and thus arises research on example selection. However, existing approaches randomly select examples or only consider the textual similarity of requirements to retrieve, leading to sub-optimal performance. In this paper, we propose a novel learning-based selection approach named LAIL (LLM-Aware In-context Learning) for code generation. Given a candidate example, we exploit LLMs themselves to estimate it by considering the generation probabilities of ground-truth programs given a requirement and the example. We then label candidate examples as positive or negative through the probability feedback. Based on the labeled data, we import a contrastive learning objective to train an effective retriever that acquires the preference of LLMs in code generation. We apply LAIL to three LLMs and evaluate it on three representative datasets (e.g., MBJP, MBPP, and MBCPP). LATA outperforms the state-of-the-art baselines by 11.58%, 6.89%, and 5.07% on CodeGen, and 4.38%, 2.85%, and 2.74% on GPT-3.5 in terms of Pass@1, respectively.

Yongjun Chen, Jia Li, Caiming Xiong

Sequential recommendation is often considered as a generative task, i.e., training a sequential encoder to generate the next item of a user's interests based on her historical interacted items. Despite their prevalence, these methods usually require training with more meaningful samples to be effective, which otherwise will lead to a poorly trained model. In this work, we propose to train the sequential recommenders as discriminators rather than generators. Instead of predicting the next item, our method trains a discriminator to distinguish if a sampled item is a 'real' target item or not. A generator, as an auxiliary model, is trained jointly with the discriminator to sample plausible alternative next items and will be thrown out after training. The trained discriminator is considered as the final SR model and denoted as \modelname. Experiments conducted on four datasets demonstrate the effectiveness and efficiency of the proposed approach.

Jia Li, Shengye Qiao, Zhirui Zhao et al.

Existing salient object detection methods often adopt deeper and wider networks for better performance, resulting in heavy computational burden and slow inference speed. This inspires us to rethink saliency detection to achieve a favorable balance between efficiency and accuracy. To this end, we design a lightweight framework while maintaining satisfying competitive accuracy. Specifically, we propose a novel trilateral decoder framework by decoupling the U-shape structure into three complementary branches, which are devised to confront the dilution of semantic context, loss of spatial structure and absence of boundary detail, respectively. Along with the fusion of three branches, the coarse segmentation results are gradually refined in structure details and boundary quality. Without adding additional learnable parameters, we further propose Scale-Adaptive Pooling Module to obtain multi-scale receptive filed. In particular, on the premise of inheriting this framework, we rethink the relationship among accuracy, parameters and speed via network depth-width tradeoff. With these insightful considerations, we comprehensively design shallower and narrower models to explore the maximum potential of lightweight SOD. Our models are purposed for different application environments: 1) a tiny version CTD-S (1.7M, 125FPS) for resource constrained devices, 2) a fast version CTD-M (12.6M, 158FPS) for speed-demanding scenarios, 3) a standard version CTD-L (26.5M, 84FPS) for high-performance platforms. Extensive experiments validate the superiority of our method, which achieves better efficiency-accuracy balance across five benchmarks.

Fang Liu, Jia Li, Li Zhang

There is growing interest in software migration as the development of software and society. Manually migrating projects between languages is error-prone and expensive. In recent years, researchers have begun to explore automatic program translation using supervised deep learning techniques by learning from large-scale parallel code corpus. However, parallel resources are scarce in the programming language domain, and it is costly to collect bilingual data manually. To address this issue, several unsupervised programming translation systems are proposed. However, these systems still rely on huge monolingual source code to train, which is very expensive. Besides, these models cannot perform well for translating the languages that are not seen during the pre-training procedure. In this paper, we propose SDA-Trans, a syntax and domain-aware model for program translation, which leverages the syntax structure and domain knowledge to enhance the cross-lingual transfer ability. SDA-Trans adopts unsupervised training on a smaller-scale corpus, including Python and Java monolingual programs. The experimental results on function translation tasks between Python, Java, and C++ show that SDA-Trans outperforms many large-scale pre-trained models, especially for unseen language translation.

Varun Akella, Razeyeh Bagherinasab, Jia Ming Li et al.

Body composition analysis is vital in assessing health conditions such as obesity, sarcopenia, and metabolic syndromes. MRI provides detailed images of skeletal muscle (SKM), visceral adipose tissue (VAT), and subcutaneous adipose tissue (SAT), but their manual segmentation is labor-intensive and limits clinical applicability. This study validates an automated tool for MRI-based 2D body composition analysis- (Data Analysis Facilitation Suite (DAFS) Express), comparing its automated measurements with expert manual segmentations using UK Biobank data. A cohort of 399 participants from the UK Biobank dataset was selected, yielding 423 single L3 slices for analysis. DAFS Express performed automated segmentations of SKM, VAT, and SAT, which were then manually corrected by expert raters for validation. Evaluation metrics included Jaccard coefficients, Dice scores, Intraclass Correlation Coefficients (ICCs), and Bland-Altman Plots to assess segmentation agreement and reliability. High agreements were observed between automated and manual segmentations with mean Jaccard scores: SKM 99.03%, VAT 95.25%, and SAT 99.57%; and mean Dice scores: SKM 99.51%, VAT 97.41%, and SAT 99.78%. Cross-sectional area comparisons showed consistent measurements with automated methods closely matching manual measurements for SKM and SAT, and slightly higher values for VAT (SKM: Auto 132.51 cm^2, Manual 132.36 cm^2; VAT: Auto 137.07 cm^2, Manual 134.46 cm^2; SAT: Auto 203.39 cm^2, Manual 202.85 cm^2). ICCs confirmed strong reliability (SKM: 0.998, VAT: 0.994, SAT: 0.994). Bland-Altman plots revealed minimal biases, and boxplots illustrated distribution similarities across SKM, VAT, and SAT areas. On average DAFS Express took 18 seconds per DICOM. This underscores its potential to streamline image analysis processes in research and clinical settings, enhancing diagnostic accuracy and efficiency.

Jia Li, Lijie Hu, Jingfeng Zhang et al.

In this paper, we address the limitations of existing text-to-image diffusion models in generating demographically fair results when given human-related descriptions. These models often struggle to disentangle the target language context from sociocultural biases, resulting in biased image generation. To overcome this challenge, we propose Fair Mapping, a flexible, model-agnostic, and lightweight approach that modifies a pre-trained text-to-image diffusion model by controlling the prompt to achieve fair image generation. One key advantage of our approach is its high efficiency. It only requires updating an additional linear network with few parameters at a low computational cost. By developing a linear network that maps conditioning embeddings into a debiased space, we enable the generation of relatively balanced demographic results based on the specified text condition. With comprehensive experiments on face image generation, we show that our method significantly improves image generation fairness with almost the same image quality compared to conventional diffusion models when prompted with descriptions related to humans. By effectively addressing the issue of implicit language bias, our method produces more fair and diverse image outputs.

Jiajin Li, Jianheng Tang, Lemin Kong et al.

In this paper, we study the design and analysis of a class of efficient algorithms for computing the Gromov-Wasserstein (GW) distance tailored to large-scale graph learning tasks. Armed with the Luo-Tseng error bound condition~\citep{luo1992error}, two proposed algorithms, called Bregman Alternating Projected Gradient (BAPG) and hybrid Bregman Proximal Gradient (hBPG) enjoy the convergence guarantees. Upon task-specific properties, our analysis further provides novel theoretical insights to guide how to select the best-fit method. As a result, we are able to provide comprehensive experiments to validate the effectiveness of our methods on a host of tasks, including graph alignment, graph partition, and shape matching. In terms of both wall-clock time and modeling performance, the proposed methods achieve state-of-the-art results.

Jia Li, Xiang Li, Xiaowei Jia et al.

Causal DAGs(Directed Acyclic Graphs) are usually considered in a 2D plane. Edges indicate causal effects' directions and imply their corresponding time-passings. Due to the natural restriction of statistical models, effect estimation is usually approximated by averaging the individuals' correlations, i.e., observational changes over a specific time. However, in the context of Machine Learning on large-scale questions with complex DAGs, such slight biases can snowball to distort global models - More importantly, it has practically impeded the development of AI, for instance, the weak generalizability of causal models. In this paper, we redefine causal DAG as \emph{do-DAG}, in which variables' values are no longer time-stamp-dependent, and timelines can be seen as axes. By geometric explanation of multi-dimensional do-DAG, we identify the \emph{Causal Representation Bias} and its necessary factors, differentiated from common confounding biases. Accordingly, a DL(Deep Learning)-based framework will be proposed as the general solution, along with a realization method and experiments to verify its feasibility.

Jia Li, Jie Cao, JunXian Duan et al.

We propose a new challenging task namely IDentity Stylization (IDS) across heterogeneous domains. IDS focuses on stylizing the content identity, rather than completely swapping it using the reference identity. We use an effective heterogeneous-network-based framework $Styleverse$ that uses a single domain-aware generator to exploit the Metaverse of diverse heterogeneous faces, based on the proposed dataset FS13 with limited data. FS13 means 13 kinds of Face Styles considering diverse lighting conditions, art representations and life dimensions. Previous similar tasks, \eg, image style transfer can handle textural style transfer based on a reference image. This task usually ignores the high structure-aware facial area and high-fidelity preservation of the content. However, Styleverse intends to controllably create topology-aware faces in the Parallel Style Universe, where the source facial identity is adaptively styled via AdaIN guided by the domain-aware and reference-aware style embeddings from heterogeneous pretrained models. We first establish the IDS quantitative benchmark as well as the qualitative Styleverse matrix. Extensive experiments demonstrate that Styleverse achieves higher-fidelity identity stylization compared with other state-of-the-art methods.

Xuefeng Jiang, Jia Li, Nannan Wu et al.

Robustness to label noise within data is a significant challenge in federated learning (FL). From the data-centric perspective, the data quality of distributed datasets can not be guaranteed since annotations of different clients contain complicated label noise of varying degrees, which causes the performance degradation. There have been some early attempts to tackle noisy labels in FL. However, there exists a lack of benchmark studies on comprehensively evaluating their practical performance under unified settings. To this end, we propose the first benchmark study FNBench to provide an experimental investigation which considers three diverse label noise patterns covering synthetic label noise, imperfect human-annotation errors and systematic errors. Our evaluation incorporates eighteen state-of-the-art methods over five image recognition datasets and one text classification dataset. Meanwhile, we provide observations to understand why noisy labels impair FL, and additionally exploit a representation-aware regularization method to enhance the robustness of existing methods against noisy labels based on our observations. Finally, we discuss the limitations of this work and propose three-fold future directions. To facilitate related communities, our source code is open-sourced at https://github.com/Sprinter1999/FNBench.

Miao Peng, Nuo Chen, Jianheng Tang et al.

Large Language Models (LLMs) have shown remarkable capabilities in knowledge-intensive tasks, while they remain vulnerable when encountering misinformation. Existing studies have explored the role of LLMs in combating misinformation, but there is still a lack of fine-grained analysis on the specific aspects and extent to which LLMs are influenced by misinformation. To bridge this gap, we present MisBench, the current largest and most comprehensive benchmark for evaluating LLMs' behavior and knowledge preference toward misinformation. MisBench consists of 10,346,712 pieces of misinformation, which uniquely considers both knowledge-based conflicts and stylistic variations in misinformation. Empirical results reveal that while LLMs demonstrate comparable abilities in discerning misinformation, they still remain susceptible to knowledge conflicts and stylistic variations. Based on these findings, we further propose a novel approach called Reconstruct to Discriminate (RtD) to strengthen LLMs' ability to detect misinformation. Our study provides valuable insights into LLMs' interactions with misinformation, and we believe MisBench can serve as an effective benchmark for evaluating LLM-based detectors and enhancing their reliability in real-world applications. Codes and data are available at https://github.com/GKNL/MisBench.

Ziyang Cheng, Zhixun Li, Yuhan Li et al.

Nowadays, real-world data, including graph-structure data, often arrives in a streaming manner, which means that learning systems need to continuously acquire new knowledge without forgetting previously learned information. Although substantial existing works attempt to address catastrophic forgetting in graph machine learning, they are all based on training from scratch with streaming data. With the rise of pretrained models, an increasing number of studies have leveraged their strong generalization ability for continual learning. Therefore, in this work, we attempt to answer whether large language models (LLMs) can mitigate catastrophic forgetting in Graph Continual Learning (GCL). We first point out that current experimental setups for GCL have significant flaws, as the evaluation stage may lead to task ID leakage. Then, we evaluate the performance of LLMs in more realistic scenarios and find that even minor modifications can lead to outstanding results. Finally, based on extensive experiments, we propose a simple-yet-effective method, Simple Graph Continual Learning (SimGCL), that surpasses the previous state-of-the-art GNN-based baseline by around 20% under the rehearsal-free constraint. To facilitate reproducibility, we have developed an easy-to-use benchmark LLM4GCL for training and evaluating existing GCL methods. The code is available at: https://github.com/ZhixunLEE/LLM4GCL.

Hongyu Li, Jia Li, Xin Ren et al.

Dust storms may remarkably degrade the imaging quality of Martian orbiters and delay the progress of mapping the global topography and geomorphology. To address this issue, this paper presents an approach that reuses the image dehazing knowledge obtained on Earth to resolve the dust-removal problem on Mars. In this approach, we collect remote-sensing images captured by Tianwen-1 and manually select hundreds of clean and dusty images. Inspired by the haze formation process on Earth, we formulate a similar visual degradation process on clean images and synthesize dusty images sharing a similar feature distribution with realistic dusty images. These realistic clean and synthetic dusty image pairs are used to train a deep model that inherently encodes dust irrelevant features and decodes them into dust-free images. Qualitative and quantitative results show that dust storms can be effectively eliminated by the proposed approach, leading to obviously improved topographical and geomorphological details of Mars.

Lixiang Zhang, Jia Li

For computational efficiency, surrogate models have been used to emulate mathematical simulators for physical or biological processes. High-speed simulation is crucial for conducting uncertainty quantification (UQ) when the simulation is repeated over many randomly sampled input points (aka, the Monte Carlo method). In some cases, UQ is only feasible with a surrogate model. Recently, Deep Neural Network (DNN) surrogate models have gained popularity for their hard-to-match emulation accuracy. However, it is well-known that DNN is prone to errors when input data are perturbed in particular ways, the very motivation for adversarial training. In the usage scenario of surrogate models, the concern is less of a deliberate attack but more of the high sensitivity of the DNN's accuracy to input directions, an issue largely ignored by researchers using emulation models. In this paper, we show the severity of this issue through empirical studies and hypothesis testing. Furthermore, we adopt methods in adversarial training to enhance the robustness of DNN surrogate models. Experiments demonstrate that our approaches significantly improve the robustness of the surrogate models without compromising emulation accuracy.

Yanlin Qi, Jia Li, Michael Zhang

Precise estimation of Crash Modification Factors (CMFs) is central to evaluating the effectiveness of various road safety treatments and prioritizing infrastructure investment accordingly. While customized study for each countermeasure scenario is desired, the conventional CMF estimation approaches rely heavily on the availability of crash data at given sites. This not only makes the estimation costly, but the results are also less transferable, since the intrinsic similarities between different safety countermeasure scenarios are not fully explored. Aiming to fill this gap, this study introduces a novel knowledge-mining framework for CMF prediction. This framework delves into the connections of existing countermeasures and reduces the reliance of CMF estimation on crash data availability and manual data collection. Specifically, it draws inspiration from human comprehension processes and introduces advanced Natural Language Processing (NLP) techniques to extract intricate variations and patterns from existing CMF knowledge. It effectively encodes unstructured countermeasure scenarios into machine-readable representations and models the complex relationships between scenarios and CMF values. This new data-driven framework provides a cost-effective and adaptable solution that complements the case-specific approaches for CMF estimation, which is particularly beneficial when availability of crash data or time imposes constraints. Experimental validation using real-world CMF Clearinghouse data demonstrates the effectiveness of this new approach, which shows significant accuracy improvements compared to baseline methods. This approach provides insights into new possibilities of harnessing accumulated transportation knowledge in various applications.

Raymond Li, Loubna Ben Allal, Yangtian Zi et al.

The BigCode community, an open-scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs), introduces StarCoder and StarCoderBase: 15.5B parameter models with 8K context length, infilling capabilities and fast large-batch inference enabled by multi-query attention. StarCoderBase is trained on 1 trillion tokens sourced from The Stack, a large collection of permissively licensed GitHub repositories with inspection tools and an opt-out process. We fine-tuned StarCoderBase on 35B Python tokens, resulting in the creation of StarCoder. We perform the most comprehensive evaluation of Code LLMs to date and show that StarCoderBase outperforms every open Code LLM that supports multiple programming languages and matches or outperforms the OpenAI code-cushman-001 model. Furthermore, StarCoder outperforms every model that is fine-tuned on Python, can be prompted to achieve 40\% pass@1 on HumanEval, and still retains its performance on other programming languages. We take several important steps towards a safe open-access model release, including an improved PII redaction pipeline and a novel attribution tracing tool, and make the StarCoder models publicly available under a more commercially viable version of the Open Responsible AI Model license.

Wenpeng Yin, Shelby Heinecke, Jia Li et al.

The distribution gap between training datasets and data encountered in production is well acknowledged. Training datasets are often constructed over a fixed period of time and by carefully curating the data to be labeled. Thus, training datasets may not contain all possible variations of data that could be encountered in real-world production environments. Tasked with building an entity resolution system - a model that identifies and consolidates data points that represent the same person - our first model exhibited a clear training-production performance gap. In this case study, we discuss our human-in-the-loop enabled, data-centric solution to closing the training-production performance divergence. We conclude with takeaways that apply to data-centric learning at large.

Dong Zhao, Jia Li, Hongyu Li et al.

Conventional CNNs-based dehazing models suffer from two essential issues: the dehazing framework (limited in interpretability) and the convolution layers (content-independent and ineffective to learn long-range dependency information). In this paper, firstly, we propose a new complementary feature enhanced framework, in which the complementary features are learned by several complementary subtasks and then together serve to boost the performance of the primary task. One of the prominent advantages of the new framework is that the purposively chosen complementary tasks can focus on learning weakly dependent complementary features, avoiding repetitive and ineffective learning of the networks. We design a new dehazing network based on such a framework. Specifically, we select the intrinsic image decomposition as the complementary tasks, where the reflectance and shading prediction subtasks are used to extract the color-wise and texture-wise complementary features. To effectively aggregate these complementary features, we propose a complementary features selection module (CFSM) to select the more useful features for image dehazing. Furthermore, we introduce a new version of vision transformer block, named Hybrid Local-Global Vision Transformer (HyLoG-ViT), and incorporate it within our dehazing networks. The HyLoG-ViT block consists of the local and the global vision transformer paths used to capture local and global dependencies. As a result, the HyLoG-ViT introduces locality in the networks and captures the global and long-range dependencies. Extensive experiments on homogeneous, non-homogeneous, and nighttime dehazing tasks reveal that the proposed dehazing network can achieve comparable or even better performance than CNNs-based dehazing models.

Beomseok Seo, Lin Lin, Jia Li

Deep neural network (DNN) models have achieved phenomenal success for applications in many domains, ranging from academic research in science and engineering to industry and business. The modeling power of DNN is believed to have come from the complexity and over-parameterization of the model, which on the other hand has been criticized for the lack of interpretation. Although certainly not true for every application, in some applications, especially in economics, social science, healthcare industry, and administrative decision making, scientists or practitioners are resistant to use predictions made by a black-box system for multiple reasons. One reason is that a major purpose of a study can be to make discoveries based upon the prediction function, e.g., to reveal the relationships between measurements. Another reason can be that the training dataset is not large enough to make researchers feel completely sure about a purely data-driven result. Being able to examine and interpret the prediction function will enable researchers to connect the result with existing knowledge or gain insights about new directions to explore. Although classic statistical models are much more explainable, their accuracy often falls considerably below DNN. In this paper, we propose an approach to fill the gap between relatively simple explainable models and DNN such that we can more flexibly tune the trade-off between interpretability and accuracy. Our main idea is a mixture of discriminative models that is trained with the guidance from a DNN. Although mixtures of discriminative models have been studied before, our way of generating the mixture is quite different.

Lixiang Zhang, Lin Lin, Jia Li

The architectures of deep neural networks (DNN) rely heavily on the underlying grid structure of variables, for instance, the lattice of pixels in an image. For general high dimensional data with variables not associated with a grid, the multi-layer perceptron and deep belief network are often used. However, it is frequently observed that those networks do not perform competitively and they are not helpful for identifying important variables. In this paper, we propose a framework that imposes on blocks of variables a chain structure obtained by step-wise greedy search so that the DNN architecture can leverage the constructed grid. We call this new neural network Deep Variable-Block Chain (DVC). Because the variable blocks are used for classification in a sequential manner, we further develop the capacity of selecting variables adaptively according to a number of regions trained by a decision tree. Our experiments show that DVC outperforms other generic DNNs and other strong classifiers. Moreover, DVC can achieve high accuracy at much reduced dimensionality and sometimes reveals drastically different sets of relevant variables for different regions.

Jia Li, Cong Fang, Zhouchen Lin

We propose a new optimization method for training feed-forward neural networks. By rewriting the activation function as an equivalent proximal operator, we approximate a feed-forward neural network by adding the proximal operators to the objective function as penalties, hence we call the lifted proximal operator machine (LPOM). LPOM is block multi-convex in all layer-wise weights and activations. This allows us to use block coordinate descent to update the layer-wise weights and activations in parallel. Most notably, we only use the mapping of the activation function itself, rather than its derivatives, thus avoiding the gradient vanishing or blow-up issues in gradient based training methods. So our method is applicable to various non-decreasing Lipschitz continuous activation functions, which can be saturating and non-differentiable. LPOM does not require more auxiliary variables than the layer-wise activations, thus using roughly the same amount of memory as stochastic gradient descent (SGD) does. We further prove the convergence of updating the layer-wise weights and activations. Experiments on MNIST and CIFAR-10 datasets testify to the advantages of LPOM.

Yu Luo, Jianbo Ye, Reginald B. Adams, et al.

Humans are arguably innately prepared to comprehend others' emotional expressions from subtle body movements. If robots or computers can be empowered with this capability, a number of robotic applications become possible. Automatically recognizing human bodily expression in unconstrained situations, however, is daunting given the incomplete understanding of the relationship between emotional expressions and body movements. The current research, as a multidisciplinary effort among computer and information sciences, psychology, and statistics, proposes a scalable and reliable crowdsourcing approach for collecting in-the-wild perceived emotion data for computers to learn to recognize body languages of humans. To accomplish this task, a large and growing annotated dataset with 9,876 video clips of body movements and 13,239 human characters, named BoLD (Body Language Dataset), has been created. Comprehensive statistical analysis of the dataset revealed many interesting insights. A system to model the emotional expressions based on bodily movements, named ARBEE (Automated Recognition of Bodily Expression of Emotion), has also been developed and evaluated. Our analysis shows the effectiveness of Laban Movement Analysis (LMA) features in characterizing arousal, and our experiments using LMA features further demonstrate computability of bodily expression. We report and compare results of several other baseline methods which were developed for action recognition based on two different modalities, body skeleton, and raw image. The dataset and findings presented in this work will likely serve as a launchpad for future discoveries in body language understanding that will enable future robots to interact and collaborate more effectively with humans.

Zihan Zhou, Siqiong He, Jia Li et al.

The capacity of automatically modeling photographic composition is valuable for many real-world machine vision applications such as digital photography, image retrieval, image understanding, and image aesthetics assessment. The triangle technique is among those indispensable composition methods on which professional photographers often rely. This paper proposes a system that can identify prominent triangle arrangements in two major categories of photographs: natural or urban scenes, and portraits. For the natural or urban scene pictures, the focus is on the effect of linear perspective. For portraits, we carefully examine the positioning of human subjects in a photo. We show that line analysis is highly advantageous for modeling composition in both categories. Based on the detected triangles, new mathematical descriptors for composition are formulated and used to retrieve similar images. Leveraging the rich source of high aesthetics photos online, similar approaches can potentially be incorporated in future smart cameras to enhance a person's photo composition skills.

Yu Zhang, James Z. Wang, Jia Li

The discrete distribution clustering algorithm, namely D2-clustering, has demonstrated its usefulness in image classification and annotation where each object is represented by a bag of weighed vectors. The high computational complexity of the algorithm, however, limits its applications to large-scale problems. We present a parallel D2-clustering algorithm with substantially improved scalability. A hierarchical structure for parallel computing is devised to achieve a balance between the individual-node computation and the integration process of the algorithm. Additionally, it is shown that even with a single CPU, the hierarchical structure results in significant speed-up. Experiments on real-world large-scale image data, Youtube video data, and protein sequence data demonstrate the efficiency and wide applicability of the parallel D2-clustering algorithm. The loss in clustering accuracy is minor in comparison with the original sequential algorithm.