Cheng Tan

Hanqun Cao, Cheng Tan, Zhangyang Gao et al.

Deep generative models have unlocked another profound realm of human creativity. By capturing and generalizing patterns within data, we have entered the epoch of all-encompassing Artificial Intelligence for General Creativity (AIGC). Notably, diffusion models, recognized as one of the paramount generative models, materialize human ideation into tangible instances across diverse domains, encompassing imagery, text, speech, biology, and healthcare. To provide advanced and comprehensive insights into diffusion, this survey comprehensively elucidates its developmental trajectory and future directions from three distinct angles: the fundamental formulation of diffusion, algorithmic enhancements, and the manifold applications of diffusion. Each layer is meticulously explored to offer a profound comprehension of its evolution. Structured and summarized approaches are presented in https://github.com/chq1155/A-Survey-on-Generative-Diffusion-Model.

Zhangyang Gao, Cheng Tan, Lirong Wu et al.

From CNN, RNN, to ViT, we have witnessed remarkable advancements in video prediction, incorporating auxiliary inputs, elaborate neural architectures, and sophisticated training strategies. We admire these progresses but are confused about the necessity: is there a simple method that can perform comparably well? This paper proposes SimVP, a simple video prediction model that is completely built upon CNN and trained by MSE loss in an end-to-end fashion. Without introducing any additional tricks and complicated strategies, we can achieve state-of-the-art performance on five benchmark datasets. Through extended experiments, we demonstrate that SimVP has strong generalization and extensibility on real-world datasets. The significant reduction of training cost makes it easier to scale to complex scenarios. We believe SimVP can serve as a solid baseline to stimulate the further development of video prediction. The code is available at \href{https://github.com/gaozhangyang/SimVP-Simpler-yet-Better-Video-Prediction}{Github}.

Siyuan Li, Zedong Wang, Zicheng Liu et al.

By contextualizing the kernel as global as possible, Modern ConvNets have shown great potential in computer vision tasks. However, recent progress on multi-order game-theoretic interaction within deep neural networks (DNNs) reveals the representation bottleneck of modern ConvNets, where the expressive interactions have not been effectively encoded with the increased kernel size. To tackle this challenge, we propose a new family of modern ConvNets, dubbed MogaNet, for discriminative visual representation learning in pure ConvNet-based models with favorable complexity-performance trade-offs. MogaNet encapsulates conceptually simple yet effective convolutions and gated aggregation into a compact module, where discriminative features are efficiently gathered and contextualized adaptively. MogaNet exhibits great scalability, impressive efficiency of parameters, and competitive performance compared to state-of-the-art ViTs and ConvNets on ImageNet and various downstream vision benchmarks, including COCO object detection, ADE20K semantic segmentation, 2D&3D human pose estimation, and video prediction. Notably, MogaNet hits 80.0% and 87.8% accuracy with 5.2M and 181M parameters on ImageNet-1K, outperforming ParC-Net and ConvNeXt-L, while saving 59% FLOPs and 17M parameters, respectively. The source code is available at https://github.com/Westlake-AI/MogaNet.

Zhangyang Gao, Cheng Tan, Pablo Chacón et al.

How can we design protein sequences folding into the desired structures effectively and efficiently? AI methods for structure-based protein design have attracted increasing attention in recent years; however, few methods can simultaneously improve the accuracy and efficiency due to the lack of expressive features and autoregressive sequence decoder. To address these issues, we propose PiFold, which contains a novel residue featurizer and PiGNN layers to generate protein sequences in a one-shot way with improved recovery. Experiments show that PiFold could achieve 51.66\% recovery on CATH 4.2, while the inference speed is 70 times faster than the autoregressive competitors. In addition, PiFold achieves 58.72\% and 60.42\% recovery scores on TS50 and TS500, respectively. We conduct comprehensive ablation studies to reveal the role of different types of protein features and model designs, inspiring further simplification and improvement. The PyTorch code is available at \href{https://github.com/A4Bio/PiFold}{GitHub}.

Cheng Tan, Siyuan Li, Zhangyang Gao et al.

Spatio-temporal predictive learning is a learning paradigm that enables models to learn spatial and temporal patterns by predicting future frames from given past frames in an unsupervised manner. Despite remarkable progress in recent years, a lack of systematic understanding persists due to the diverse settings, complex implementation, and difficult reproducibility. Without standardization, comparisons can be unfair and insights inconclusive. To address this dilemma, we propose OpenSTL, a comprehensive benchmark for spatio-temporal predictive learning that categorizes prevalent approaches into recurrent-based and recurrent-free models. OpenSTL provides a modular and extensible framework implementing various state-of-the-art methods. We conduct standard evaluations on datasets across various domains, including synthetic moving object trajectory, human motion, driving scenes, traffic flow and weather forecasting. Based on our observations, we provide a detailed analysis of how model architecture and dataset properties affect spatio-temporal predictive learning performance. Surprisingly, we find that recurrent-free models achieve a good balance between efficiency and performance than recurrent models. Thus, we further extend the common MetaFormers to boost recurrent-free spatial-temporal predictive learning. We open-source the code and models at https://github.com/chengtan9907/OpenSTL.

Zicheng Liu, Siyuan Li, Ge Wang et al.

Mixup is an efficient data augmentation approach that improves the generalization of neural networks by smoothing the decision boundary with mixed data. Recently, dynamic mixup methods have improved previous static policies effectively (e.g., linear interpolation) by maximizing target-related salient regions in mixed samples, but excessive additional time costs are not acceptable. These additional computational overheads mainly come from optimizing the mixed samples according to the mixed labels. However, we found that the extra optimizing step may be redundant because label-mismatched mixed samples are informative hard mixed samples for deep models to localize discriminative features. In this paper, we thus are not trying to propose a more complicated dynamic mixup policy but rather an efficient mixup objective function with a decoupled regularizer named Decoupled Mixup (DM). The primary effect is that DM can adaptively utilize those hard mixed samples to mine discriminative features without losing the original smoothness of mixup. As a result, DM enables static mixup methods to achieve comparable or even exceed the performance of dynamic methods without any extra computation. This also leads to an interesting objective design problem for mixup training that we need to focus on both smoothing the decision boundaries and identifying discriminative features. Extensive experiments on supervised and semi-supervised learning benchmarks across seven datasets validate the effectiveness of DM as a plug-and-play module. Source code and models are available at https://github.com/Westlake-AI/openmixup

Cheng Tan, Jingxuan Wei, Zhangyang Gao et al.

Multimodal reasoning is a challenging task that requires models to reason across multiple modalities to answer questions. Existing approaches have made progress by incorporating language and visual modalities into a two-stage reasoning framework, separating rationale generation from answer inference. However, these approaches often fall short due to the inadequate quality of the generated rationales. In this work, we delve into the importance of rationales in model reasoning. We observe that when rationales are completely accurate, the model's accuracy significantly improves, highlighting the need for high-quality rationale generation. Motivated by this, we propose MC-CoT, a self-consistency training strategy that generates multiple rationales and answers, subsequently selecting the most accurate through a voting process. This approach not only enhances the quality of generated rationales but also leads to more accurate and robust answers. Through extensive experiments, we demonstrate that our approach significantly improves model performance across various benchmarks. Remarkably, we show that even smaller base models, when equipped with our proposed approach, can achieve results comparable to those of larger models, illustrating the potential of our approach in harnessing the power of rationales for improved multimodal reasoning. The code is available at https://github.com/chengtan9907/mc-cot.

Cheng Tan, Yijie Zhang, Zhangyang Gao et al.

While artificial intelligence has made remarkable strides in revealing the relationship between biological macromolecules' primary sequence and tertiary structure, designing RNA sequences based on specified tertiary structures remains challenging. Though existing approaches in protein design have thoroughly explored structure-to-sequence dependencies in proteins, RNA design still confronts difficulties due to structural complexity and data scarcity. Moreover, direct transplantation of protein design methodologies into RNA design fails to achieve satisfactory outcomes although sharing similar structural components. In this study, we aim to systematically construct a data-driven RNA design pipeline. We crafted a large, well-curated benchmark dataset and designed a comprehensive structural modeling approach to represent the complex RNA tertiary structure. More importantly, we proposed a hierarchical data-efficient representation learning framework that learns structural representations through contrastive learning at both cluster-level and sample-level to fully leverage the limited data. By constraining data representations within a limited hyperspherical space, the intrinsic relationships between data points could be explicitly imposed. Moreover, we incorporated extracted secondary structures with base pairs as prior knowledge to facilitate the RNA design process. Extensive experiments demonstrate the effectiveness of our proposed method, providing a reliable baseline for future RNA design tasks. The source code and benchmark dataset are available at https://github.com/A4Bio/RDesign.

Cheng Tan, Zhangyang Gao, Lirong Wu et al.

Spatiotemporal predictive learning aims to generate future frames by learning from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal predictive learning, in which the spatial encoder and decoder capture intra-frame features and the middle temporal module catches inter-frame correlations. While the mainstream methods employ recurrent units to capture long-term temporal dependencies, they suffer from low computational efficiency due to their unparallelizable architectures. To parallelize the temporal module, we propose the Temporal Attention Unit (TAU), which decomposes the temporal attention into intra-frame statical attention and inter-frame dynamical attention. Moreover, while the mean squared error loss focuses on intra-frame errors, we introduce a novel differential divergence regularization to take inter-frame variations into account. Extensive experiments demonstrate that the proposed method enables the derived model to achieve competitive performance on various spatiotemporal prediction benchmarks.

Jiawei Liu, Jinkun Lin, Fabian Ruffy et al.

Deep-learning (DL) compilers such as TVM and TensorRT are increasingly being used to optimize deep neural network (DNN) models to meet performance, resource utilization and other requirements. Bugs in these compilers can result in models whose semantics differ from the original ones, producing incorrect results that corrupt the correctness of downstream applications. However, finding bugs in these compilers is challenging due to their complexity. In this work, we propose a new fuzz testing approach for finding bugs in deep-learning compilers. Our core approach consists of (i) generating diverse yet valid DNN test models that can exercise a large part of the compiler's transformation logic using light-weight operator specifications; (ii) performing gradient-based search to find model inputs that avoid any floating-point exceptional values during model execution, reducing the chance of missed bugs or false alarms; and (iii) using differential testing to identify bugs. We implemented this approach in NNSmith which has found 72 new bugs for TVM, TensorRT, ONNXRuntime, and PyTorch to date. Of these 58 have been confirmed and 51 have been fixed by their respective project maintainers.

Jiangbin Zheng, Yile Wang, Cheng Tan et al.

Sign language recognition (SLR) is a weakly supervised task that annotates sign videos as textual glosses. Recent studies show that insufficient training caused by the lack of large-scale available sign datasets becomes the main bottleneck for SLR. Most SLR works thereby adopt pretrained visual modules and develop two mainstream solutions. The multi-stream architectures extend multi-cue visual features, yielding the current SOTA performances but requiring complex designs and might introduce potential noise. Alternatively, the advanced single-cue SLR frameworks using explicit cross-modal alignment between visual and textual modalities are simple and effective, potentially competitive with the multi-cue framework. In this work, we propose a novel contrastive visual-textual transformation for SLR, CVT-SLR, to fully explore the pretrained knowledge of both the visual and language modalities. Based on the single-cue cross-modal alignment framework, we propose a variational autoencoder (VAE) for pretrained contextual knowledge while introducing the complete pretrained language module. The VAE implicitly aligns visual and textual modalities while benefiting from pretrained contextual knowledge as the traditional contextual module. Meanwhile, a contrastive cross-modal alignment algorithm is designed to explicitly enhance the consistency constraints. Extensive experiments on public datasets (PHOENIX-2014 and PHOENIX-2014T) demonstrate that our proposed CVT-SLR consistently outperforms existing single-cue methods and even outperforms SOTA multi-cue methods.

Cheng Tan, Zhangyang Gao, Hanqun Cao et al.

The secondary structure of ribonucleic acid (RNA) is more stable and accessible in the cell than its tertiary structure, making it essential for functional prediction. Although deep learning has shown promising results in this field, current methods suffer from poor generalization and high complexity. In this work, we reformulate the RNA secondary structure prediction as a K-Rook problem, thereby simplifying the prediction process into probabilistic matching within a finite solution space. Building on this innovative perspective, we introduce RFold, a simple yet effective method that learns to predict the most matching K-Rook solution from the given sequence. RFold employs a bi-dimensional optimization strategy that decomposes the probabilistic matching problem into row-wise and column-wise components to reduce the matching complexity, simplifying the solving process while guaranteeing the validity of the output. Extensive experiments demonstrate that RFold achieves competitive performance and about eight times faster inference efficiency than the state-of-the-art approaches. The code and Colab demo are available in (http://github.com/A4Bio/RFold).

Bozhen Hu, Jun Xia, Jiangbin Zheng et al.

The prediction of protein structures from sequences is an important task for function prediction, drug design, and related biological processes understanding. Recent advances have proved the power of language models (LMs) in processing the protein sequence databases, which inherit the advantages of attention networks and capture useful information in learning representations for proteins. The past two years have witnessed remarkable success in tertiary protein structure prediction (PSP), including evolution-based and single-sequence-based PSP. It seems that instead of using energy-based models and sampling procedures, protein language model (pLM)-based pipelines have emerged as mainstream paradigms in PSP. Despite the fruitful progress, the PSP community needs a systematic and up-to-date survey to help bridge the gap between LMs in the natural language processing (NLP) and PSP domains and introduce their methodologies, advancements and practical applications. To this end, in this paper, we first introduce the similarities between protein and human languages that allow LMs extended to pLMs, and applied to protein databases. Then, we systematically review recent advances in LMs and pLMs from the perspectives of network architectures, pre-training strategies, applications, and commonly-used protein databases. Next, different types of methods for PSP are discussed, particularly how the pLM-based architectures function in the process of protein folding. Finally, we identify challenges faced by the PSP community and foresee promising research directions along with the advances of pLMs. This survey aims to be a hands-on guide for researchers to understand PSP methods, develop pLMs and tackle challenging problems in this field for practical purposes.

Tong Geng, Chunshu Wu, Yongan Zhang et al.

Graph Convolutional Networks (GCNs) have drawn tremendous attention in the past three years. Compared with other deep learning modalities, high-performance hardware acceleration of GCNs is as critical but even more challenging. The hurdles arise from the poor data locality and redundant computation due to the large size, high sparsity, and irregular non-zero distribution of real-world graphs. In this paper we propose a novel hardware accelerator for GCN inference, called I-GCN, that significantly improves data locality and reduces unnecessary computation. The mechanism is a new online graph restructuring algorithm we refer to as islandization. The proposed algorithm finds clusters of nodes with strong internal but weak external connections. The islandization process yields two major benefits. First, by processing islands rather than individual nodes, there is better on-chip data reuse and fewer off-chip memory accesses. Second, there is less redundant computation as aggregation for common/shared neighbors in an island can be reused. The parallel search, identification, and leverage of graph islands are all handled purely in hardware at runtime working in an incremental pipeline. This is done without any preprocessing of the graph data or adjustment of the GCN model structure. Experimental results show that I-GCN can significantly reduce off-chip accesses and prune 38% of aggregation operations, leading to performance speedups over CPUs, GPUs, the prior art GCN accelerators of 5549x, 403x, and 5.7x on average, respectively.

Cheng Tan, Zhangyang Gao, Lirong Wu et al.

Recent advances in contrastive learning have enlightened diverse applications across various semi-supervised fields. Jointly training supervised learning and unsupervised learning with a shared feature encoder becomes a common scheme. Though it benefits from taking advantage of both feature-dependent information from self-supervised learning and label-dependent information from supervised learning, this scheme remains suffering from bias of the classifier. In this work, we systematically explore the relationship between self-supervised learning and supervised learning, and study how self-supervised learning helps robust data-efficient deep learning. We propose hyperspherical consistency regularization (HCR), a simple yet effective plug-and-play method, to regularize the classifier using feature-dependent information and thus avoid bias from labels. Specifically, HCR first projects logits from the classifier and feature projections from the projection head on the respective hypersphere, then it enforces data points on hyperspheres to have similar structures by minimizing binary cross entropy of pairwise distances' similarity metrics. Extensive experiments on semi-supervised and weakly-supervised learning demonstrate the effectiveness of our method, by showing superior performance with HCR.

Zhangyang Gao, Cheng Tan, Stan Z. Li

Have you ever been troubled by the complexity and computational cost of SE(3) protein structure modeling and been amazed by the simplicity and power of language modeling? Recent work has shown promise in simplifying protein structures as sequences of protein angles; therefore, language models could be used for unconstrained protein backbone generation. Unfortunately, such simplification is unsuitable for the constrained protein inpainting problem, where the model needs to recover masked structures conditioned on unmasked ones, as it dramatically increases the computing cost of geometric constraints. To overcome this dilemma, we suggest inserting a hidden \textbf{a}tomic \textbf{d}irection \textbf{s}pace (\textbf{ADS}) upon the language model, converting invariant backbone angles into equivalent direction vectors and preserving the simplicity, called Seq2Direct encoder ($\text{Enc}_{s2d}$). Geometric constraints could be efficiently imposed on the newly introduced direction space. A Direct2Seq decoder ($\text{Dec}_{d2s}$) with mathematical guarantees is also introduced to develop a \textbf{SDS} ($\text{Enc}_{s2d}$+$\text{Dec}_{d2s}$) model. We apply the SDS model as the denoising neural network during the conditional diffusion process, resulting in a constrained generative model--\textbf{DiffSDS}. Extensive experiments show that the plug-and-play ADS could transform the language model into a strong structural model without loss of simplicity. More importantly, the proposed DiffSDS outperforms previous strong baselines by a large margin on the task of protein inpainting.

Zhangyang Gao, Cheng Tan, Lirong Wu et al.

Can we inject the pocket-ligand interaction knowledge into the pre-trained model and jointly learn their chemical space? Pretraining molecules and proteins has attracted considerable attention in recent years, while most of these approaches focus on learning one of the chemical spaces and lack the injection of biological knowledge. We propose a co-supervised pretraining (CoSP) framework to simultaneously learn 3D pocket and ligand representations. We use a gated geometric message passing layer to model both 3D pockets and ligands, where each node's chemical features, geometric position and orientation are considered. To learn biological meaningful embeddings, we inject the pocket-ligand interaction knowledge into the pretraining model via contrastive loss. Considering the specificity of molecules, we further propose a chemical similarity-enhanced negative sampling strategy to improve the contrastive learning performance. Through extensive experiments, we conclude that CoSP can achieve competitive results in pocket matching, molecule property predictions, and virtual screening.

Zhangyang Gao, Yuqi Hu, Cheng Tan et al.

Is there a unified model for generating molecules considering different conditions, such as binding pockets and chemical properties? Although target-aware generative models have made significant advances in drug design, they do not consider chemistry conditions and cannot guarantee the desired chemical properties. Unfortunately, merging the target-aware and chemical-aware models into a unified model to meet customized requirements may lead to the problem of negative transfer. Inspired by the success of multi-task learning in the NLP area, we use prefix embeddings to provide a novel generative model that considers both the targeted pocket's circumstances and a variety of chemical properties. All conditional information is represented as learnable features, which the generative model subsequently employs as a contextual prompt. Experiments show that our model exhibits good controllability in both single and multi-conditional molecular generation. The controllability enables us to outperform previous structure-based drug design methods. More interestingly, we open up the attention mechanism and reveal coupling relationships between conditions, providing guidance for multi-conditional molecule generation.

Cheng Tan, Zhangyang Gao, Jun Xia et al.

The linear sequence of amino acids determines protein structure and function. Protein design, known as the inverse of protein structure prediction, aims to obtain a novel protein sequence that will fold into the defined structure. Recent works on computational protein design have studied designing sequences for the desired backbone structure with local positional information and achieved competitive performance. However, similar local environments in different backbone structures may result in different amino acids, indicating that protein structure's global context matters. Thus, we propose the Global-Context Aware generative de novo protein design method (GCA), consisting of local and global modules. While local modules focus on relationships between neighbor amino acids, global modules explicitly capture non-local contexts. Experimental results demonstrate that the proposed GCA method outperforms state-of-the-arts on de novo protein design. Our code and pretrained model will be released.

Cheng Tan, Zhangyang Gao, Siyuan Li et al.

Recent years have witnessed remarkable advances in spatiotemporal predictive learning, with methods incorporating auxiliary inputs, complex neural architectures, and sophisticated training strategies. While SimVP has introduced a simpler, CNN-based baseline for this task, it still relies on heavy Unet-like architectures for spatial and temporal modeling, which still suffers from high complexity and computational overhead. In this paper, we propose SimVPv2, a streamlined model that eliminates the need for Unet architectures and demonstrates that plain stacks of convolutional layers, enhanced with an efficient Gated Spatiotemporal Attention mechanism, can deliver state-of-the-art performance. SimVPv2 not only simplifies the model architecture but also improves both performance and computational efficiency. On the standard Moving MNIST benchmark, SimVPv2 achieves superior performance compared to SimVP, with fewer FLOPs, about half the training time, and 60% faster inference efficiency. Extensive experiments across eight diverse datasets, including real-world tasks such as traffic forecasting and climate prediction, further demonstrate that SimVPv2 offers a powerful yet straightforward solution, achieving robust generalization across various spatiotemporal learning scenarios. We believe the proposed SimVPv2 can serve as a solid baseline to benefit the spatiotemporal predictive learning community.

Jiangbin Zheng, Siyuan Li, Cheng Tan et al.

Sign Language (SL), as the mother tongue of the deaf community, is a special visual language that most hearing people cannot understand. In recent years, neural Sign Language Translation (SLT), as a possible way for bridging communication gap between the deaf and the hearing people, has attracted widespread academic attention. We found that the current mainstream end-to-end neural SLT models, which tries to learning language knowledge in a weakly supervised manner, could not mine enough semantic information under the condition of low data resources. Therefore, we propose to introduce additional word-level semantic knowledge of sign language linguistics to assist in improving current end-to-end neural SLT models. Concretely, we propose a novel neural SLT model with multi-modal feature fusion based on the dynamic graph, in which the cross-modal information, i.e. text and video, is first assembled as a dynamic graph according to their correlation, and then the graph is processed by a multi-modal graph encoder to generate the multi-modal embeddings for further usage in the subsequent neural translation models. To the best of our knowledge, we are the first to introduce graph neural networks, for fusing multi-modal information, into neural sign language translation models. Moreover, we conducted experiments on a publicly available popular SLT dataset RWTH-PHOENIX-Weather-2014T. and the quantitative experiments show that our method can improve the model.

Zhe Li, Zhangyang Gao, Cheng Tan et al.

The pre-training architectures of large language models encompass various types, including autoencoding models, autoregressive models, and encoder-decoder models. We posit that any modality can potentially benefit from a large language model, as long as it undergoes vector quantization to become discrete tokens. Inspired by GLM, we propose a General Point Model (GPM) which seamlessly integrates autoencoding and autoregressive tasks in point cloud transformer. This model is versatile, allowing fine-tuning for downstream point cloud representation tasks, as well as unconditional and conditional generation tasks. GPM enhances masked prediction in autoencoding through various forms of mask padding tasks, leading to improved performance in point cloud understanding. Additionally, GPM demonstrates highly competitive results in unconditional point cloud generation tasks, even exhibiting the potential for conditional generation tasks by modifying the input's conditional information. Compared to models like Point-BERT, MaskPoint and PointMAE, our GPM achieves superior performance in point cloud understanding tasks. Furthermore, the integration of autoregressive and autoencoding within the same transformer underscores its versatility across different downstream tasks.

Bozhen Hu, Bin Gao, Cheng Tan et al.

Defect detection plays a crucial role in infrared non-destructive testing systems, offering non-contact, safe, and efficient inspection capabilities. However, challenges such as low resolution, high noise, and uneven heating in infrared thermal images hinder comprehensive and accurate defect detection. In this study, we propose DefectSAM, a novel approach for segmenting defects on highly noisy thermal images based on the widely adopted model, Segment Anything (SAM)\cite{kirillov2023segany}. Harnessing the power of a meticulously curated dataset generated through labor-intensive lab experiments and valuable prompts from experienced experts, DefectSAM surpasses existing state-of-the-art segmentation algorithms and achieves significant improvements in defect detection rates. Notably, DefectSAM excels in detecting weaker and smaller defects on complex and irregular surfaces, reducing the occurrence of missed detections and providing more accurate defect size estimations. Experimental studies conducted on various materials have validated the effectiveness of our solutions in defect detection, which hold significant potential to expedite the evolution of defect detection tools, enabling enhanced inspection capabilities and accuracy in defect identification.

Lirong Wu, Haitao Lin, Guojiang Zhao et al.

Recent years have witnessed great success in handling graph-related tasks with Graph Neural Networks (GNNs). However, most existing GNNs are based on message passing to perform feature aggregation and transformation, where the structural information is explicitly involved in the forward propagation by coupling with node features through graph convolution at each layer. As a result, subtle feature noise or structure perturbation may cause severe error propagation, resulting in extremely poor robustness. In this paper, we rethink the roles played by graph structural information in graph data training and identify that message passing is not the only path to modeling structural information. Inspired by this, we propose a simple but effective Graph Structure Self-Contrasting (GSSC) framework that learns graph structural information without message passing. The proposed framework is based purely on Multi-Layer Perceptrons (MLPs), where the structural information is only implicitly incorporated as prior knowledge to guide the computation of supervision signals, substituting the explicit message propagation as in GNNs. Specifically, it first applies structural sparsification to remove potentially uninformative or noisy edges in the neighborhood, and then performs structural self-contrasting in the sparsified neighborhood to learn robust node representations. Finally, structural sparsification and self-contrasting are formulated as a bi-level optimization problem and solved in a unified framework. Extensive experiments have qualitatively and quantitatively demonstrated that the GSSC framework can produce truly encouraging performance with better generalization and robustness than other leading competitors.

Jessica Maghakian, Paul Mineiro, Kishan Panaganti et al.

In an era of countless content offerings, recommender systems alleviate information overload by providing users with personalized content suggestions. Due to the scarcity of explicit user feedback, modern recommender systems typically optimize for the same fixed combination of implicit feedback signals across all users. However, this approach disregards a growing body of work highlighting that (i) implicit signals can be used by users in diverse ways, signaling anything from satisfaction to active dislike, and (ii) different users communicate preferences in different ways. We propose applying the recent Interaction Grounded Learning (IGL) paradigm to address the challenge of learning representations of diverse user communication modalities. Rather than requiring a fixed, human-designed reward function, IGL is able to learn personalized reward functions for different users and then optimize directly for the latent user satisfaction. We demonstrate the success of IGL with experiments using simulations as well as with real-world production traces.

Cheng Tan, Jue Wang, Zhangyang Gao et al.

Spatio-temporal predictive learning plays a crucial role in self-supervised learning, with wide-ranging applications across a diverse range of fields. Previous approaches for temporal modeling fall into two categories: recurrent-based and recurrent-free methods. The former, while meticulously processing frames one by one, neglect short-term spatio-temporal information redundancies, leading to inefficiencies. The latter naively stack frames sequentially, overlooking the inherent temporal dependencies. In this paper, we re-examine the two dominant temporal modeling approaches within the realm of spatio-temporal predictive learning, offering a unified perspective. Building upon this analysis, we introduce USTEP (Unified Spatio-TEmporal Predictive learning), an innovative framework that reconciles the recurrent-based and recurrent-free methods by integrating both micro-temporal and macro-temporal scales. Extensive experiments on a wide range of spatio-temporal predictive learning demonstrate that USTEP achieves significant improvements over existing temporal modeling approaches, thereby establishing it as a robust solution for a wide range of spatio-temporal applications.

Siyuan Li, Luyuan Zhang, Zedong Wang et al.

As the deep learning revolution marches on, self-supervised learning has garnered increasing attention in recent years thanks to its remarkable representation learning ability and the low dependence on labeled data. Among these varied self-supervised techniques, masked modeling has emerged as a distinctive approach that involves predicting parts of the original data that are proportionally masked during training. This paradigm enables deep models to learn robust representations and has demonstrated exceptional performance in the context of computer vision, natural language processing, and other modalities. In this survey, we present a comprehensive review of the masked modeling framework and its methodology. We elaborate on the details of techniques within masked modeling, including diverse masking strategies, recovering targets, network architectures, and more. Then, we systematically investigate its wide-ranging applications across domains. Furthermore, we also explore the commonalities and differences between masked modeling methods in different fields. Toward the end of this paper, we conclude by discussing the limitations of current techniques and point out several potential avenues for advancing masked modeling research. A paper list project with this survey is available at \url{https://github.com/Lupin1998/Awesome-MIM}.

Cheng Tan, Dongxin Lyu, Siyuan Li et al.

Large Language Models (LLMs) have demonstrated wide-ranging applications across various fields and have shown significant potential in the academic peer-review process. However, existing applications are primarily limited to static review generation based on submitted papers, which fail to capture the dynamic and iterative nature of real-world peer reviews. In this paper, we reformulate the peer-review process as a multi-turn, long-context dialogue, incorporating distinct roles for authors, reviewers, and decision makers. We construct a comprehensive dataset containing over 26,841 papers with 92,017 reviews collected from multiple sources, including the top-tier conference and prestigious journal. This dataset is meticulously designed to facilitate the applications of LLMs for multi-turn dialogues, effectively simulating the complete peer-review process. Furthermore, we propose a series of metrics to evaluate the performance of LLMs for each role under this reformulated peer-review setting, ensuring fair and comprehensive evaluations. We believe this work provides a promising perspective on enhancing the LLM-driven peer-review process by incorporating dynamic, role-based interactions. It aligns closely with the iterative and interactive nature of real-world academic peer review, offering a robust foundation for future research and development in this area. We open-source the dataset at https://github.com/chengtan9907/ReviewMT.

Lirong Wu, Haitao Lin, Zhangyang Gao et al.

Deep learning on graphs has recently achieved remarkable success on a variety of tasks, while such success relies heavily on the massive and carefully labeled data. However, precise annotations are generally very expensive and time-consuming. To address this problem, self-supervised learning (SSL) is emerging as a new paradigm for extracting informative knowledge through well-designed pretext tasks without relying on manual labels. In this survey, we extend the concept of SSL, which first emerged in the fields of computer vision and natural language processing, to present a timely and comprehensive review of existing SSL techniques for graph data. Specifically, we divide existing graph SSL methods into three categories: contrastive, generative, and predictive. More importantly, unlike other surveys that only provide a high-level description of published research, we present an additional mathematical summary of existing works in a unified framework. Furthermore, to facilitate methodological development and empirical comparisons, we also summarize the commonly used datasets, evaluation metrics, downstream tasks, open-source implementations, and experimental study of various algorithms. Finally, we discuss the technical challenges and potential future directions for improving graph self-supervised learning. Latest advances in graph SSL are summarized in a GitHub repository https://github.com/LirongWu/awesome-graph-self-supervised-learning.

Zhe Li, Laurence T. Yang, Bocheng Ren et al.

The scarcity of annotated data has sparked significant interest in unsupervised pre-training methods that leverage medical reports as auxiliary signals for medical visual representation learning. However, existing research overlooks the multi-granularity nature of medical visual representation and lacks suitable contrastive learning techniques to improve the models' generalizability across different granularities, leading to the underutilization of image-text information. To address this, we propose MLIP, a novel framework leveraging domain-specific medical knowledge as guiding signals to integrate language information into the visual domain through image-text contrastive learning. Our model includes global contrastive learning with our designed divergence encoder, local token-knowledge-patch alignment contrastive learning, and knowledge-guided category-level contrastive learning with expert knowledge. Experimental evaluations reveal the efficacy of our model in enhancing transfer performance for tasks such as image classification, object detection, and semantic segmentation. Notably, MLIP surpasses state-of-the-art methods even with limited annotated data, highlighting the potential of multimodal pre-training in advancing medical representation learning.

Siyuan Li, Zedong Wang, Zicheng Liu et al.

Similar to natural language models, pre-trained genome language models are proposed to capture the underlying intricacies within genomes with unsupervised sequence modeling. They have become essential tools for researchers and practitioners in biology. However, the hand-crafted tokenization policies used in these models may not encode the most discriminative patterns from the limited vocabulary of genomic data. In this paper, we introduce VQDNA, a general-purpose framework that renovates genome tokenization from the perspective of genome vocabulary learning. By leveraging vector-quantized codebooks as learnable vocabulary, VQDNA can adaptively tokenize genomes into pattern-aware embeddings in an end-to-end manner. To further push its limits, we propose Hierarchical Residual Quantization (HRQ), where varying scales of codebooks are designed in a hierarchy to enrich the genome vocabulary in a coarse-to-fine manner. Extensive experiments on 32 genome datasets demonstrate VQDNA's superiority and favorable parameter efficiency compared to existing genome language models. Notably, empirical analysis of SARS-CoV-2 mutations reveals the fine-grained pattern awareness and biological significance of learned HRQ vocabulary, highlighting its untapped potential for broader applications in genomics.

Ming Hu, Chenglong Ma, Wei Li et al. · pku

Scientific Large Language Models (Sci-LLMs) are transforming how knowledge is represented, integrated, and applied in scientific research, yet their progress is shaped by the complex nature of scientific data. This survey presents a comprehensive, data-centric synthesis that reframes the development of Sci-LLMs as a co-evolution between models and their underlying data substrate. We formulate a unified taxonomy of scientific data and a hierarchical model of scientific knowledge, emphasizing the multimodal, cross-scale, and domain-specific challenges that differentiate scientific corpora from general natural language processing datasets. We systematically review recent Sci-LLMs, from general-purpose foundations to specialized models across diverse scientific disciplines, alongside an extensive analysis of over 270 pre-/post-training datasets, showing why Sci-LLMs pose distinct demands -- heterogeneous, multi-scale, uncertainty-laden corpora that require representations preserving domain invariance and enabling cross-modal reasoning. On evaluation, we examine over 190 benchmark datasets and trace a shift from static exams toward process- and discovery-oriented assessments with advanced evaluation protocols. These data-centric analyses highlight persistent issues in scientific data development and discuss emerging solutions involving semi-automated annotation pipelines and expert validation. Finally, we outline a paradigm shift toward closed-loop systems where autonomous agents based on Sci-LLMs actively experiment, validate, and contribute to a living, evolving knowledge base. Collectively, this work provides a roadmap for building trustworthy, continually evolving artificial intelligence (AI) systems that function as a true partner in accelerating scientific discovery.

Yijie Zhang, Zhangyang Gao, Cheng Tan et al.

Predicting protein stability changes induced by single-point mutations has been a persistent challenge over the years, attracting immense interest from numerous researchers. The ability to precisely predict protein thermostability is pivotal for various subfields and applications in biochemistry, including drug development, protein evolution analysis, and enzyme synthesis. Despite the proposition of multiple methodologies aimed at addressing this issue, few approaches have successfully achieved optimal performance coupled with high computational efficiency. Two principal hurdles contribute to the existing challenges in this domain. The first is the complexity of extracting and aggregating sufficiently representative features from proteins. The second refers to the limited availability of experimental data for protein mutation analysis, further complicating the comprehensive evaluation of model performance on unseen data samples. With the advent of Large Language Models(LLM), such as the ESM models in protein research, profound interpretation of protein features is now accessibly aided by enormous training data. Therefore, LLMs are indeed to facilitate a wide range of protein research. In our study, we introduce an ESM-assisted efficient approach that integrates protein sequence and structural features to predict the thermostability changes in protein upon single-point mutations. Furthermore, we have curated a dataset meticulously designed to preclude data leakage, corresponding to two extensively employed test datasets, to facilitate a more equitable model comparison.

Bozhen Hu, Zelin Zang, Cheng Tan et al.

Protein representation learning is critical in various tasks in biology, such as drug design and protein structure or function prediction, which has primarily benefited from protein language models and graph neural networks. These models can capture intrinsic patterns from protein sequences and structures through masking and task-related losses. However, the learned protein representations are usually not well optimized, leading to performance degradation due to limited data, difficulty adapting to new tasks, etc. To address this, we propose a new \underline{d}eep \underline{m}anifold \underline{t}ransformation approach for universal \underline{p}rotein \underline{r}epresentation \underline{l}earning (DMTPRL). It employs manifold learning strategies to improve the quality and adaptability of the learned embeddings. Specifically, we apply a novel manifold learning loss during training based on the graph inter-node similarity. Our proposed DMTPRL method outperforms state-of-the-art baselines on diverse downstream tasks across popular datasets. This validates our approach for learning universal and robust protein representations. We promise to release the code after acceptance.

Zhangyang Gao, Xingran Chen, Cheng Tan et al.

Is there a unified framework for graph-based retrosynthesis prediction? Through analysis of full-, semi-, and non-template retrosynthesis methods, we discovered that they strive to strike an optimal balance between combinability and consistency: \textit{Should atoms be combined as motifs to simplify the molecular editing process, or should motifs be broken down into atoms to reduce the vocabulary and improve predictive consistency?} Recent works have studied several specific cases, while none of them explores different combinability-consistency trade-offs. Therefore, we propose MotifRetro, a dynamic motif editing framework for retrosynthesis prediction that can explore the entire trade-off space and unify graph-based models. MotifRetro comprises two components: RetroBPE, which controls the combinability-consistency trade-off, and a motif editing model, where we introduce a novel LG-EGAT module to dynamiclly add motifs to the molecule. We conduct extensive experiments on USPTO-50K to explore how the trade-off affects the model performance and finally achieve state-of-the-art performance.

Zhangyang Gao, Cheng Tan, Stan Z. Li

Recent studies have shown competitive performance in protein design that aims to find the amino acid sequence folding into the desired structure. However, most of them disregard the importance of predictive confidence, fail to cover the vast protein space, and do not incorporate common protein knowledge. After witnessing the great success of pretrained models on diverse protein-related tasks and the fact that recovery is highly correlated with confidence, we wonder whether this knowledge can push the limits of protein design further. As a solution, we propose a knowledge-aware module that refines low-quality residues. We also introduce a memory-retrieval mechanism to save more than 50\% of the training time. We extensively evaluate our proposed method on the CATH, TS50, and TS500 datasets and our results show that our Knowledge-Design method outperforms the previous PiFold method by approximately 9\% on the CATH dataset. Specifically, Knowledge-Design is the first method that achieves 60+\% recovery on CATH, TS50 and TS500 benchmarks. We also provide additional analysis to demonstrate the effectiveness of our proposed method. The code will be publicly available.

Zhangyang Gao, Cheng Tan, Lirong Wu et al.

Recently, template-based (TB) and template-free (TF) molecule graph learning methods have shown promising results to retrosynthesis. TB methods are more accurate using pre-encoded reaction templates, and TF methods are more scalable by decomposing retrosynthesis into subproblems, i.e., center identification and synthon completion. To combine both advantages of TB and TF, we suggest breaking a full-template into several semi-templates and embedding them into the two-step TF framework. Since many semi-templates are reduplicative, the template redundancy can be reduced while the essential chemical knowledge is still preserved to facilitate synthon completion. We call our method SemiRetro, introduce a new GNN layer (DRGAT) to enhance center identification, and propose a novel self-correcting module to improve semi-template classification. Experimental results show that SemiRetro significantly outperforms both existing TB and TF methods. In scalability, SemiRetro covers 98.9\% data using 150 semi-templates, while previous template-based GLN requires 11,647 templates to cover 93.3\% data. In top-1 accuracy, SemiRetro exceeds template-free G2G 4.8\% (class known) and 6.0\% (class unknown). Besides, SemiRetro has better training efficiency than existing methods.

Cheng Tan, Zhangyang Gao, Stan Z. Li

Generating molecules with desired biological activities has attracted growing attention in drug discovery. Previous molecular generation models are designed as chemocentric methods that hardly consider the drug-target interaction, limiting their practical applications. In this paper, we aim to generate molecular drugs in a target-aware manner that bridges biological activity and molecular design. To solve this problem, we compile a benchmark dataset from several publicly available datasets and build baselines in a unified framework. Building on the recent advantages of flow-based molecular generation models, we propose SiamFlow, which forces the flow to fit the distribution of target sequence embeddings in latent space. Specifically, we employ an alignment loss and a uniform loss to bring target sequence embeddings and drug graph embeddings into agreements while avoiding collapse. Furthermore, we formulate the alignment into a one-to-many problem by learning spaces of target sequence embeddings. Experiments quantitatively show that our proposed method learns meaningful representations in the latent space toward the target-aware molecular graph generation and provides an alternative approach to bridge biology and chemistry in drug discovery.

Zhangyang Gao, Cheng Tan, Stan Z. Li

While DeepMind has tentatively solved protein folding, its inverse problem -- protein design which predicts protein sequences from their 3D structures -- still faces significant challenges. Particularly, the lack of large-scale standardized benchmark and poor accuray hinder the research progress. In order to standardize comparisons and draw more research interest, we use AlphaFold DB, one of the world's largest protein structure databases, to establish a new graph-based benchmark -- AlphaDesign. Based on AlphaDesign, we propose a new method called ADesign to improve accuracy by introducing protein angles as new features, using a simplified graph transformer encoder (SGT), and proposing a confidence-aware protein decoder (CPD). Meanwhile, SGT and CPD also improve model efficiency by simplifying the training and testing procedures. Experiments show that ADesign significantly outperforms previous graph models, e.g., the average accuracy is improved by 8\%, and the inference speed is 40+ times faster than before.

Heting Liu, Zhichao Li, Cheng Tan et al.

Graphics processing units (GPUs) are the de facto standard for processing deep learning (DL) tasks. Meanwhile, GPU failures, which are inevitable, cause severe consequences in DL tasks: they disrupt distributed trainings, crash inference services, and result in service level agreement violations. To mitigate the problem caused by GPU failures, we propose to predict failures by using ML models. This paper is the first to study prediction models of GPU failures under large-scale production deep learning workloads. As a starting point, we evaluate classic prediction models and observe that predictions of these models are both inaccurate and unstable. To improve the precision and stability of predictions, we propose several techniques, including parallel and cascade model-ensemble mechanisms and a sliding training method. We evaluate the performances of our various techniques on a four-month production dataset including 350 million entries. The results show that our proposed techniques improve the prediction precision from 46.3\% to 84.0\%.