Graph contrastive learning (GCL) has emerged as a dominant technique for graph representation which maximizes the mutual information between paired augmentations that share same semantics. Unfortunately, it is difficult to preserve semantics well during in view of diverse nature data. Currently, data GCL broadly fall into three unsatisfactory ways. First, can be manually picked per dataset by trial-and-errors. Second, selected via cumbersome search. Third, obtained with expensive domain...

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, prediction model completely built upon CNN trained by MSE loss an end-to-end fashion. Without introducing any additional tricks complicated strategies,...

Deep learning on graphs has recently achieved remarkable success a variety of tasks, while such relies heavily the massive and carefully labeled data. However, precise annotations are generally very expensive time-consuming. To address this problem, self-supervised (SSL) is emerging as new paradigm for extracting informative knowledge through well-designed pretext tasks without relying manual labels. In survey, we extend concept SSL, which first emerged in fields computer vision natural...

Spatiotemporal predictive learning aims to generate future frames by from historical frames. In this paper, we investigate existing methods and present a general framework of spatiotemporal learning, in which the spatial encoder decoder capture intra-frame features middle temporal module catches inter-frame correlations. While mainstream employ recurrent units long-term dependencies, they suffer low computational efficiency due their unparallelizable architectures. To parallelize module,...

Noisy labels, resulting from mistakes in manual labeling or webly data collecting for supervised learning, can cause neural networks to overfit the misleading information and degrade generalization performance. Self-supervised learning works absence of labels thus eliminates negative impact noisy labels. Motivated by co-training with both view self-supervised view, we propose a simple yet effective method called Co-learning performs cooperative way. The constraints intrinsic similarity...

Spatio-temporal forecasting is challenging attributing to the high nonlinearity in temporal dynamics as well complex location-characterized patterns spatial domains, especially fields like weather forecasting. Graph convolutions are usually used for modeling dependency meteorology handle irregular distribution of sensors' location. In this work, a novel graph-based convolution imitating meteorological flows proposed capture local patterns. Based on assumption smoothness patterns, we propose...

Geometric deep learning has recently achieved great success in non-Euclidean domains, and on 3D structures of large biomolecules is emerging as a distinct research area. However, its efficacy largely constrained due to the limited quantity structural data. Meanwhile, protein language models trained substantial 1D sequences have shown burgeoning capabilities with scale broad range applications. Several preceding studies consider combining these different modalities promote representation...

Spatio-temporal predictive learning is a paradigm that enables models to learn spatial and temporal patterns by predicting future frames from given past in an unsupervised manner. Despite remarkable progress recent years, lack of systematic understanding persists due the diverse settings, complex implementation, difficult reproducibility. Without standardization, comparisons can be unfair insights inconclusive. To address this dilemma, we propose OpenSTL, comprehensive benchmark for...

Antibodies are Y-shaped proteins that protect the host by binding to specific antigens, and their is mainly determined Complementary Determining Regions (CDRs) in antibody. Despite great progress made CDR design, existing computational methods still encounter several challenges: 1) poor capability of modeling complex CDRs with long sequences due insufficient contextual information; 2) conditioned on pre-given antigenic epitopes static interaction target antibody; 3) neglect specificity...

The method of importance map has been widely adopted in DNN-based lossy image compression to achieve bit allocation according the contents. However, insufficient bits non-important regions often leads severe distortion at low bpp (bits per pixel), which hampers development efficient content-weighted systems. This paper rethinks content-based by using Generative Adversarial Network (GAN) reconstruct regions. Moreover, multiscale pyramid decomposition is applied both encoder and discriminator...

Graph neural networks (GNNs) have been playing important roles in various graph-related tasks. However, most existing GNNs are based on the assumption of homophily, so they cannot be directly generalized to heterophily settings where connected nodes may different features and class labels. Moreover, real-world graphs often arise from highly entangled latent factors, but tend ignore this simply denote heterogeneous relations between as binary-valued homogeneous edges. In article, we propose a...

Recent years have witnessed great success in handling graph-related tasks with Graph Neural Networks (GNNs). Despite their <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">academic</i> success, Multi-Layer Perceptrons (MLPs) remain the primary workhorse for practical xmlns:xlink="http://www.w3.org/1999/xlink">industrial</i> applications. One reason such an academic-industry gap is neighborhood-fetching latency incurred by data dependency GNNs....

Compound-Protein Interaction (CPI) prediction aims to predict the pattern and strength of compound-protein interactions for rational drug discovery. Existing deep learning-based methods utilize only single modality protein sequences or structures lack co-modeling joint distribution two modalities, which may lead significant performance drops in complex real-world scenarios due various factors, e.g., missing domain shifting. More importantly, these model at a fixed scale, neglecting more...

The proteins that exist today have been optimized over billions of years natural evolution, during which nature creates random mutations and selects them. discovery functionally promising is challenged by the limited evolutionary accessible regions, i.e., only a small region on fitness landscape beneficial. There numerous priors used to constrain protein evolution regions landscapes with high-fitness variants, among change in binding free energy (DDG) complexes upon one most commonly priors....

Graph edge perturbations are dedicated to damaging the prediction of graph neural networks by modifying structure. Previous gray-box attackers employ gradients from surrogate model locate vulnerable edges perturb However, unreliability exists in on structures, which is rarely studied previous works. In this paper, we discuss and analyze errors caused structural gradients. These arise rough gradient usage due discreteness structure meta-gradient order address these problems, propose a novel...

Graph Neural Network (GNN) has emerged as a predominant tool for graph data analysis. Despite their proliferation, the low-quality labels of many real-world graphs will undermine performance dramatically. Existing studies on learning neural networks with noisy mainly focus independent and thus cannot fully exploit structural information data. Currently, there are few robustness to graph-structured even if this problem is commonly seen in settings. To remedy deficiency, we propose <italic...

Antibodies are crucial proteins produced by the immune system in response to foreign substances or antigens. The specificity of an antibody is determined its complementarity-determining regions (CDRs), which located variable domains chains and form antigen-binding site. Previous studies have utilized complex techniques generate CDRs, but they suffer from inadequate geometric modeling. Moreover, common iterative refinement strategies lead inefficient inference. In this paper, we propose a...

The data storage has been one of the bottlenecks in surveillance systems. conventional video compression schemes such as H.264 and H.265 do not fully utilize low information density characteristic video, they attach equal importance to foreground background when performing compression. In this article, we propose a novel scheme that compresses separately. ratio is greatly improved by sharing among adjacent frames through an adaptive updating interpolation module. Besides, present two...

Recent advances in contrastive learning have enlightened diverse applications across various semi-supervised fields. Jointly training supervised and unsupervised with a shared feature encoder becomes common scheme. Though it benefits from taking advantage of both feature-dependent information self-supervised label-dependent learning, this scheme remains suffering bias the classifier. In work, we systematically explore relationship between study how helps robust data-efficient deep learning....

Graph neural networks (GNNs) have recently achieved remarkable success on a variety of graph-related tasks, while such relies heavily given graph structure that may not always be available in real-world applications. To address this problem, learning (GSL) is emerging as promising research topic where task-specific and GNN parameters are jointly learned an end-to-end unified framework. Despite their great progress, existing approaches mostly focus the design similarity metrics or...

Abstract Therapeutic peptides have proven to great pharmaceutical value and potential in recent decades. However, methods of AI-assisted peptide drug discovery are not fully explored. To fill the gap, we propose a target-aware design method called PPF low , based on conditional flow matching torus manifolds, model internal geometries torsion angles for structure design. Besides, establish protein-peptide binding dataset named PPBench2024 void massive data task structure-based allow training...

Homophily principle, \ie{} nodes with the same labels or similar attributes are more likely to be connected, has been commonly believed main reason for superiority of Graph Neural Networks (GNNs) over traditional (NNs) on graph-structured data, especially node-level tasks. However, recent work identified a non-trivial set datasets where GNN's performance compared NN's is not satisfactory. Heterophily, i.e. low homophily, considered cause this empirical observation. People have begun revisit...