Predicting the performance of deep learning (DL) models, such as execution time and resource utilization, is crucial for Neural Architecture Search (NAS), DL cluster schedulers, and other technologies that advance deep learning. The representation of a model is the foundation for its performance prediction. However, existing methods cannot comprehensively represent diverse model configurations, resulting in unsatisfactory accuracy. To address this, we represent a model as a graph that includes the topology, along with node, edge, and global features, all of which are crucial for effectively capturing the performance of the model. Based on this representation, we propose PerfSeer, a novel predictor that uses a Graph Neural Network (GNN)-based performance prediction model, SeerNet. SeerNet fully leverages the topology and various features, while incorporating optimizations such as Synergistic Max-Mean aggregation (SynMM) and Global-Node Perspective Boost (GNPB) to more effectively capture the critical performance information, enabling it to predict the performance of models accurately. Furthermore, SeerNet can be extended to SeerNet-Multi by using Project Conflicting Gradients (PCGrad), enabling efficient simultaneous prediction of multiple performance metrics without significantly affecting accuracy. We constructed a dataset containing performance metrics for 53k+ model configurations, including execution time, memory usage, and Streaming Multiprocessor (SM) utilization during both training and inference. The evaluation results show that PerfSeer outperforms nn-Meter, Brp-NAS, and DIPPM.
论文检索
输入标题、作者或关键词，从 2,893 篇学术成果中精准定位
QiMeng-TensorOp: One-Line Prompt is Enough for High-Performance Tensor Operator Generation with Hardware Primitives
PDF ↗Computation-intensive tensor operators constitute over 90% of the computations in Large Language Models (LLMs) and Deep Neural Networks. Automatically and efficiently generating high-performance tensor operators with hardware primitives is crucial for diverse and ever-evolving hardware architectures like RISC-V, ARM, and GPUs, as manually optimized implementation takes at least months and lacks portability. LLMs excel at generating high-level language codes, but they struggle to fully comprehend hardware characteristics and produce high-performance tensor operators. We introduce a tensor-operator auto-generation framework with a one-line user prompt (QiMeng-TensorOp), which enables LLMs to automatically exploit hardware characteristics to generate tensor operators with hardware primitives, and tune parameters for optimal performance across diverse hardware. Experimental results on various hardware platforms, SOTA LLMs, and typical tensor operators demonstrate that QiMeng-TensorOp effectively unleashes the computing capability of various hardware platforms, and automatically generates tensor operators of superior performance. Compared with vanilla LLMs, QiMeng-TensorOp achieves up to 1291× performance improvement. Even compared with human experts, QiMeng-TensorOp could reach 251% of OpenBLAS on RISC-V CPUs, and 124% of cuBLAS on NVIDIA GPUs. Additionally, QiMeng-TensorOp also significantly reduces development costs by 200× compared with human experts.
Federated graph learning is a widely recognized technique that promotes collaborative training of graph neural networks (GNNs) by multi-client graphs.However, existing approaches heavily rely on the communication of model parameters or gradients for federated optimization and fail to adequately address the data heterogeneity introduced by intricate and diverse graph distributions. Although some methods attempt to share additional messages among the server and clients to improve federated convergence during communication, they introduce significant privacy risks and increase communication overhead. To address these issues, we introduce the concept of a condensed graph as a novel optimization carrier to address FGL data heterogeneity and propose a new FGL paradigm called FedGM. Specifically, we utilize a generalized condensation graph consensus to aggregate comprehensive knowledge from distributed graphs, while minimizing communication costs and privacy risks through a single transmission of the condensed data. Extensive experiments on six public datasets consistently demonstrate the superiority of FedGM over state-of-the-art baselines, highlighting its potential for a novel FGL paradigm.
Empowering Vision Transformers with Multi-Scale Causal Intervention for Long-Tailed Image Classification
PDF ↗Causal inference has emerged as a promising approach to mitigate long-tail classification by handling the biases introduced by class imbalance. However, along with the change of advanced backbone models from Convolutional Neural Networks (CNNs) to Visual Transformers (ViT), existing causal models may not achieve an expected performance gain. This paper investigates the influence of existing causal models on CNNs and ViT variants, highlighting that ViT's global feature representation makes it hard for causal methods to model associations between fine-grained features and predictions, which leads to difficulties in classifying tail classes with similar visual appearance. To address these issues, this paper proposes TSCNet, a two-stage causal modeling method to discover fine-grained causal associations through multi-scale causal interventions. Specifically, in the hierarchical causal representation learning stage (HCRL), it decouples the background and objects, applying backdoor interventions at both the patch and feature level to prevent model from using class-irrelevant areas to infer labels which enhances fine-grained causal representation. In the counterfactual logits' bias calibration stage (CLBC), it refines the optimization of model's decision boundary by adaptive constructing counterfactual balanced data distribution to remove the spurious associations in the logits caused by data distribution. Extensive experiments conducted on various long-tail benchmarks demonstrate that the proposed TSCNet can eliminate multiple biases introduced by data imbalance, which outperforms existing methods.
A Priori Estimation of the Approximation, Optimization and Generalization Errors of Random Neural Networks for Solving Partial Differential Equations
PDF ↗In recent years, neural networks have achieved remarkable progress in various fields and have also drawn much attention in applying them on scientific problems. A line of methods involving neural networks for solving partial differential equations (PDEs), such as Physics-Informed Neural Networks (PINNs) and the Deep Ritz Method (DRM), has emerged. Although these methods outperform classical numerical methods in certain cases, the optimization problems involving neural networks are typically non-convex and non-smooth, which can result in unsatisfactory solutions for PDEs. In contrast to deterministic neural networks, the hidden weights of random neural networks are sampled from some prior distribution and only the output weights participate in training. This makes training much simpler, but it remains unclear how to select the prior distribution. In this paper, we focus on Barron type functions and approximate them under Sobolev norms by random neural networks with clear prior distribution. In addition to the approximation error, we also derive bounds for the optimization and generalization errors of random neural networks for solving PDEs when the solutions are Barron type functions.
Decision trees and forests have achieved successes in various real applications, most working with all testing classes known in training data. In this work, we focus on learning with augmented class via forests, where an augmented class may appear in testing data yet not in training data. We incorporate information of augmented class into trees' splitting, that is, augmented Gini impurity, a new splitting criterion is introduced to exploit some unlabeled data from testing distribution. We then develop the Learning with Augmented Class via Forests (short for LACForest) approach, which constructs shallow forests according to the augmented Gini impurity and then splits forests with pseudo-labeled augmented instances for better performance. We also develop deep neural forests via an optimization objective based on our augmented Gini impurity, which essentially utilizes the representation power of neural networks for forests. Theoretically, we present the convergence analysis for our augmented Gini impurity, and we finally conduct experiments to evaluate our approaches. The code is available at https://github.com/nju-xuf/LACForest.
Adaptive gradient methods, primarily based on Adam, are prevalent in training neural networks, adjusting step sizes via exponentially decaying averages of gradients and squared gradients. Adam assigns small weights to distant gradients, termed long-tail gradients in this paper. However, these gradients persistently influence update behavior, potentially degrading generalization performance. To address this issue, we incorporate a restart mechanism into moment estimations, proposing AdaR (ADAptive gradient methods via Restarting moment estimations). Specifically, AdaR divides a training epoch into fixed-iteration intervals, alternating between two sets of moment estimations for parameter updates and discarding prior moment estimations at the beginning of each interval. Within each interval, one set updates parameters and will be discarded in the subsequent interval, while the other is reset at the midpoint to estimate moments for updates in the subsequent interval. The restart mechanism cyclically discards distant gradients, initiates fresh moment estimations for parameter updates, and stabilizes training. By prioritizing recent gradients, the method increases estimation accuracy and enhances step size adjustment. Empirically, AdaR outperforms state-of-the-art optimization algorithms on image classification and language modeling tasks, demonstrating superior generalization and faster convergence.
Spiking neural network (SNN) is an event-driven neural network that can greatly reduce the power consumption of the conventional artificial neural networks (ANN). Many ANN models can be converted to SNN models when the activation function is ReLU. For ANN models with other activation functions, such as the Leaky ReLU function, the converted SNN models either suffer from serious accuracy degradation or require a long time step. In this paper, we propose a fast and accurate ANN-SNN conversion algorithm for models with the Leaky ReLU function. We design a novel neuron model that supports negative spikes. To address the problem of long tail distribution in the activation values, we propose a threshold optimization algorithm based on the variance of the activation values. To avoid the problem of error accumulation, we jointly calibrate all layers in the SNN model with adaptive weighting. Experiment results verify the effectiveness of the proposed algorithm.
DANCE: Resource-Efficient Neural Architecture Search with Data-Aware and Continuous Adaptation
PDF ↗Neural Architecture Search (NAS) has emerged as a powerful approach for automating neural network design. However, existing NAS methods face critical limitations in real-world deployments: architectures lack adaptability across scenarios, each deployment context requires costly separate searches, and performance consistency across diverse platforms remains challenging. We propose DANCE (Dynamic Architectures with Neural Continuous Evolution), which reformulates architecture search as a continuous evolution problem through learning distributions over architectural components. DANCE introduces three key innovations: a continuous architecture distribution enabling smooth adaptation, a unified architecture space with learned selection gates for efficient sampling, and a multi-stage training strategy for effective deployment optimization. Extensive experiments across five datasets demonstrate DANCE's effectiveness. Our method consistently outperforms state-of-the-art NAS approaches in terms of accuracy while significantly reducing search costs. Under varying computational constraints, DANCE maintains robust performance while smoothly adapting architectures to different hardware requirements. The code and appendix can be found at https://github.com/Applied-Machine-Learning-Lab/DANCE.
This paper presents HGEN that pioneers ensemble learning for heterogeneous graphs. We argue that the heterogeneity in node types, nodal features, and local neighborhood topology poses significant challenges for ensemble learning, particularly in accommodating diverse graph learners. Our HGEN framework ensembles multiple learners through a meta-path and transformation-based optimization pipeline to uplift classification accuracy. Specifically, HGEN uses meta-path combined with random dropping to create Allele Graph Neural Networks (GNNs), whereby the base graph learners are trained and aligned for later ensembling. To ensure effective ensemble learning, HGEN presents two key components:1) a residual-attention mechanism to calibrate allele GNNs of different meta-paths, thereby enforcing node embeddings to focus on more informative graphs to improve base learner accuracy, and 2) a correlation-regularization term to enlarge the disparity among embedding matrices generated from different meta-paths, thereby enriching base learner diversity. We analyze the convergence of HGEN and attest its higher regularization magnitude over simple voting. Experiments on five heterogeneous networks validate that HGEN consistently outperforms its state-of-the-art competitors by substantial margin. Codes are available at https://github.com/Chrisshen12/HGEN.
Graph Neural Networks (GNNs), with their ability to effectively handle non-Euclidean data structures, have demonstrated state-of-the-art performance in learning node and graph-level representations. However, GNNs face significant computational overhead due to their message-passing mechanisms, making them impractical for real-time large-scale applications. Recently, Graph-to-MLP (G2M) knowledge distillation has emerged as a promising solution, utilizing MLPs to reduce inference latency. However, existing methods often lack structural awareness (SA), limiting their ability to capture essential graph-specific information. Moreover, some methods require access to large-scale graphs, undermining their scalability. To address these issues, we propose SALE-MLP (Structure-Aware Latent Embeddings for GNN-to-Graph-Free MLP Distillation), a novel graph-free and structure-aware approach that leverages unsupervised structural losses to align the MLP feature space with the underlying graph structure. SALE-MLP does not rely on precomputed GNN embeddings nor require graph during inference, making it efficient for real-world applications. Extensive experiments demonstrate that SALE-MLP outperforms existing G2M methods across tasks and datasets, achieving 3–4% improvement in node classification for inductive settings while maintaining strong transductive performance.
Mitigating Over-Smoothing in Graph Neural Networks via Separation Coefficient-Guided Adaptive Graph Structure Adjustment
PDF ↗As the number of layers in Graph Neural Networks (GNNs) increases, over-smoothing becomes more severe, causing intra-class feature distances to shrink, while heterogeneous representations tend to converge. Most existing methods attempt to address this issue by employing heuristic shortcut mechanisms or optimizing objectives to constrain inter-class feature differences. However, these approaches fail to establish a theoretical connection between message passing and the variation in inter-class feature differences, making it challenging to design methods that target the key influencing factors. To address this gap, this paper first introduces the concept of the separation coefficient, which quantifies the contraction of feature distances between classes during multi-layer message passing. Based on this theory, we propose a low-complexity, pluggable, pseudo-label-based adaptive graph structure adjustment method. This approach effectively enhances the separation coefficient of inter-class features while maintaining intra-class compactness, thereby alleviating the convergence of heterogeneous representations caused by multi-layer aggregation. Experimental results demonstrate that the proposed method significantly improves the discriminability of node representations and enhances node classification performance across various datasets and foundational models.
Integrating multi-omics datasets through data-driven analysis offers a comprehensive understanding of the complex biological processes underlying various diseases, particularly cancer. Graph Neural Networks (GNNs) have recently demonstrated remarkable ability to exploit relational structures in biological data, enabling advances in multi-omics integration for cancer subtype classification. Existing approaches often neglect the intricate coupling between heterogeneous omics, limiting their capacity to resolve subtle cancer subtype heterogeneity critical for precision oncology. To address these limitations, we propose a framework named Graph Transformer for Multi-omics Cancer Subtype Classification (GTMancer). This framework builds upon the GNN optimization problem and extends its application to complex multi-omics data. Specifically, our method leverages contrastive learning to embed multi-omics data into a unified semantic space. We unroll the multiplex graph optimization problem in that unified space and introduce dual sets of attention coefficients to capture structural graph priors both within and among multi-omics data. This approach enables global omics information to guide the refining of the representations of individual omics. Empirical experiments on seven real-world cancer datasets demonstrate that GTMancer outperforms existing state-of-the-art algorithms.
Recently, the mathematical tool from fractal geometry (i.e., fractal dimension) has been employed to investigate optimization trajectory-dependent generalization ability for some pointwise learning models with independent and identically distributed (i.i.d.) observations. This paper goes beyond the limitations of pointwise learning and i.i.d. samples, and establishes generalization bounds for pairwise learning with uniformly strong mixing samples. The derived theoretical results fill the gap of trajectory-dependent generalization analysis for pairwise learning, and can be applied to wide learning paradigms, e.g., metric learning, ranking and gradient learning. Technically, our framework brings concentration estimation with Rademacher complexity and trajectory-dependent fractal dimension together in a coherent way for felicitous learning theory analysis. In addition, the efficient computation of fractal dimension can be guaranteed for random algorithms (e.g., stochastic gradient descent algorithm for deep neural networks) by bridging topological data analysis tools and the trajectory-dependent fractal dimension.
Multivariate time series (MTS) forecasting endeavors to anticipate the forthcoming sequence of interdependent variables through the utilization of past observations. The prevailing methodologies, relying on deep neural networks, Transformer, or information bottleneck frameworks, persist in confronting challenges such as overlooking or inadequately capturing the inter / intra-series correlations evident in practical MTS datasets. In response to these challenges, we introduce a conditional information bottleneck-based strategy for MTS forecasting, grounded in information theory. Initially, we establish a conditional information bottleneck principle to capture the inter-series correlations via conditioning on non-target variables. Subsequently, a conditional mutual information-based technique is introduced to extract intra-series correlations by conditioning historical data, ensuring temporal consistency within each variable. Lastly, we devise a unified optimization objective and propose a training algorithm to collectively capture inter / intra-series correlations. Empirical investigations on authentic datasets underscore the superiority of our proposed approach over other cutting-edge competitors. Our code is available at https: //github.com/Xinhui-Lee/CIB-MTSF.
Mixed precision quantization (MPQ) is an effective quantization approach to achieve accuracy-complexity trade-off of neural network, through assigning different bit-widths to network activations and weights in each layer. The typical way of existing MPQ methods is to optimize quantization policies (i.e., bit-width allocation) in a gradient descent manner, termed as Differentiable MPQ (DMPQ). At the end of the search, the bit-width associated to the quantization parameters which has the largest value will be selected to form the final mixed precision quantization policy, with the implicit assumption that the values of quantization parameters reflect the operation contribution to the accuracy improvement. While much has been discussed about the MPQ’s improvement, the bit-width selection process has received little attention. We study this problem and argue that the magnitude of quantization parameters does not necessarily reflect the actual contribution of the bit-width to the task performance. Then, we propose a Shapley-based MPQ (SMPQ) method, which measures the bit-width operation’s direct contribution on the MPQ task. To reduce computation cost, a Monte Carlo sampling-based approximation strategy is proposed for Shapley computation. Extensive experiments on mainstream benchmarks demonstrate that our SMPQ consistently achieves state-of-the-art performance than gradient-based competitors.
The combination of Spiking Neural Networks (SNNs) with Vision Transformer architectures has attracted significant attention due to the great potential for energy-efficient and high-performance computing paradigms. However, a substantial performance gap still exists between SNN-based and ANN-based transformer architectures. While existing methods propose spiking self-attention mechanisms that are successfully combined with SNNs, the overall architectures proposed by these methods suffer from a bottleneck in effectively extracting features from different image scales. In this paper, we address this issue and propose MSVIT, a novel spike-driven Transformer architecture, which firstly uses multi-scale spiking attention (MSSA) to enrich the capability of spiking attention blocks. We validate our approach across various main data sets. The experimental results indicate that our MSVIT outperforms existing SNN-based models, positioning itself as a state-of-the-art solution among NN-transformer architectures. The codes are available at https://github.com/Nanhu-AI-Lab/MSViT.
Model-Based Closed-Loop Control Algorithm for Stochastic Partial Differential Equation Control
PDF ↗Neural operators have demonstrated promise in modeling and controlling systems governed by Partial Differential Equations (PDEs). Beyond PDEs, Stochastic Partial Differential Equations (SPDEs) play a critical role in modeling systems influenced by randomness, with applications in finance, physics, and beyond. However, controlling SPDE-governed systems remains a significant challenge. On the one hand, the regularity of the system's state (which can be intuitively understood as smoothness) deteriorates, making modeling and generalization more challenging. On the other hand, this stochasticity also renders control more unstable and thus less accurate. To address this gap, we propose the Model-Based Closed-Loop Control Algorithm (MB-CC), the first model-based closed-loop control method for SPDEs. MB-CC introduces two key innovations to enhance control robustness and efficiency: a Regularity Feature (RF) block and a closed-loop strategy with an operator-encoded policy network. The RF block, inspired by the regularity structure theory of SPDEs, addresses noise-induced irregularities by transforming the network's input—including the system state and noise-perturbed external forces—into a refined feature space for improved forward prediction. Compared to previous works using regularity features, we introduce a new parameterization, data augmentation, and extend the RF block as a plug-and-play component. Additionally, to achieve closed-loop control, we introduce an operator-encoded policy network to map the current state to optimal control, which integrates physical priors and swiftly makes decisions based on states returned by the environment. We conduct a systematic evaluation of MB-CC on two notable SPDEs, showcasing its effectiveness and efficiency. The ablation studies show its ability to handle stochasticity more effectively.
Personalized federated graph neural networks (PFGNN) are an emerging technology that allows multiple graph data owners to collaboratively train personalized models without sharing raw data. However, the Non-IID nature of graph data can cause the coupling of global and local knowledge parameters, which disrupts the optimization in personalized federated learning. Additionally, node neighbors may carry global and local knowledge, and their direct inclusion in training may introduce noise, degrading federated model performance. In this work, we propose the Adaptive Decoupling Personalized Federated Graph Neural Network (ADPFedGNN), which leverages multi-party collaboration to train personalized models for classifying local client graph nodes. We use two automatically updated masks and mutual information minimization to decouple global and local parameters in FGNN. We employ reinforcement learning to adaptively select appropriate neighbors for training global or local knowledge-related parameters while filtering out irrelevant nodes. We also design a personalized federated masked parameter aggregation mechanism that efficiently updates local personalized model parameters and aggregates the masked parameters. Experimental results on three public datasets demonstrate that ADPFedGNN outperforms existing methods, achieving average improvements of 5.66 percent, 5.83 percent, and 12.45 percent in ACC, F1, and Recall, respectively.
Neural architecture search-based multi-modal classification (NAS-MMC) aims to automatically find optimal network structures for improving the multi-modal classification performance. However, most current NAS-MMC methods are quite time-consuming during the training process. In this paper, we propose a knowledge sharing-based neural architecture search (KS-NAS) method for multi-modal classification. The KS-NAS optimizes the search process by introducing a dynamically updated knowledge base to reduce the consumption of computational resource. Specifically, during the deep evolutionary search, individuals in the initial population acquire initial parameters from a knowledge base, and then undergo training and optimization until convergence is reached, avoiding the need for training from scratch. The knowledge base is dynamically updated by aggregating the parameters of high-quality individuals trained within the population, thus progressively improving the quality of the knowledge base. As the population evolves, the knowledge base continues to optimize, ensuring that subsequent individuals can obtain higher-quality initialization parameters, which significantly accelerates the training speed of the population. Experimental results show that the KS-NAS method achieves state-of-the-art results in terms of classification performance and training efficiency across multiple popular multi-modal tasks.