With the rapid growth and rising complexity of web content, edge-deployed LLMs have become essential for enhancing users' online experiences. However, sparsely-activated Mixture-of-Experts (MoE) models, which are well-suited for edge scenarios, face significant memory bottleneck challenges. Offload-based methods have been proposed to mitigate the problem, but they face difficulties with expert prediction. To promote the application of MoE models in edge scenarios, we propose Fate, an offloading system designed for MoE models to enable efficient inference in resource-constrained environments. The key insight behind Fate is that gate inputs from adjacent layers can be effectively used for expert prefetching, achieving high prediction accuracy. Furthermore, Fate employs a shallow-favoring expert caching strategy that increases the expert hit rate to 99%. Additionally, Fate integrates tailored quantization strategies for cache optimization and I/O efficiency. Experimental results show that, compared to baselines, Fate achieves up to 1.34×-5.07× prefill speedup and 1.26×-4.41× decoding speedup, while maintaining inference quality.

The rapid growth of deep learning models and the increasing demand for large-scale datasets have posed unprece dented challenges for data loading and memory management. Existing frameworks (e.g., PyTorch, TensorFlow) often encounter performance bottlenecks when handling large datasets resulting in inefficiencies and excessive memory usage. To address these issues, we propose Lafa, a dynamic metadata loading mechanism optimized for efficient large-scale dataset processing. Lafa introduces the .Lafa format and an adaptive loading strategy with three modes to balance memory usage and loading performance, along with a local shuffle approach that reduces memory overhead and computational complexity while preserving data randomness. Experimental results on GPU (RTX 3090) and Ascend (910A) platforms demonstrate that Lafa significantly improves memory efficiency compared to existing frameworks. Specifically, for every 10 million samples loaded, Lafa reduces additional memory consumption by a factor of 1.33× to 31.34× across various dataset types, relative to the most memory-efficient baseline among PyTorch, TensorFlow, and MindSpore.

Mixture of Experts (MoE), with its distinctive sparse structure, enables the scaling of language models up to trillions of parameters without significantly increasing computational costs. However, the substantial parameter size presents a challenge for inference, as the expansion in GPU memory cannot keep pace with the growth in parameters. Although offloading techniques utilise memory from the CPU and disk and parallelise the I/O and computation for efficiency, the computation for each expert in MoE models is often less than the I/O, resulting in numerous bubbles in the pipeline.

Understanding and reasoning about dynamics governed by physical laws through visual observation, akin to human capabilities in the real world, poses significant challenges. Currently, object-centric dynamic simulation methods, which emulate human behavior, have achieved notable progress but overlook two critical aspects: 1) the integration of physical knowledge into models. Humans gain physical insights by observing the world and apply this knowledge to accurately reason about various dynamic scenarios; 2) the validation of model adaptability across diverse scenarios. Real-world dynamics, especially those involving fluids and objects, demand models that not only capture object interactions but also simulate fluid flow characteristics. To address these gaps, we introduce SlotPi, a slot-based physics-informed object-centric reasoning model. SlotPi integrates a physical module based on Hamiltonian principles with a spatio-temporal prediction module for dynamic forecasting. Our experiments highlight the model's strengths in tasks such as prediction and Visual Question Answering (VQA) on benchmark and fluid datasets. Furthermore, we have created a real-world dataset encompassing object interactions, fluid dynamics, and fluid-object interactions, on which we validated our model's capabilities. The model's robust performance across all datasets underscores its strong adaptability, laying a foundation for developing more advanced world models.

Simulation of fluid flows is crucial for modeling physical phenomena like meteorology, aerodynamics, and biomedicine. Classical numerical solvers often require fine spatiotemporal grids to satisfy stability, consistency, and convergence conditions, leading to substantial computational costs. Although machine learning has demonstrated better efficiency, they typically suffer from issues of interpretability, generalizability, and data dependency. Hence, we propose a learnable and differentiable finite volume solver, called LDSolver, designed for efficient and accurate simulation of fluid flows on spatiotemporal coarse grids. LDSolver comprises two key components: (1) a differentiable finite volume solver, and (2) an learnable module providing equivalent approximation for fluxes (derivatives and interpolations), and temporal error correction on coarse grids. Even with limited training data (e.g., only a few trajectories), our model could accelerate the simulation while maintaining a high accuracy with superior generalizability. Experiments on different flow systems (e.g., Burgers, decaying, forced and shear flows) show that LDSolver achieves state-of-the-art performance, surpassing baseline models with notable margins.

Graph-based fraud detection has widespread application in modern industry scenarios, such as spam review and malicious account detection. While considerable efforts have been devoted to designing adequate fraud detectors, the interpretability of their results has often been overlooked. Previous works have attempted to generate explanations for specific instances using post-hoc explaining methods such as a GNNExplainer. However, post-hoc explanations can not facilitate the model predictions and the computational cost of these methods cannot meet practical requirements, thus limiting their application in real-world scenarios. To address these issues, we propose SEFraud, a novel graph-based self-explainable fraud detection framework that simultaneously tackles fraud detection and result in interpretability. Concretely, SEFraud first leverages customized heterogeneous graph transformer networks with learnable feature masks and edge masks to learn expressive representations from the informative heterogeneously typed transactions. A new triplet loss is further designed to enhance the performance of mask learning. Empirical results on various datasets demonstrate the effectiveness of SEFraud as it shows considerable advantages in both the fraud detection performance and interpretability of prediction results. Specifically, SEFraud achieves the most significant improvement with 8.6% on AUC and 8.5% on Recall over the second best on fraud detection, as well as an average of 10x speed-up regarding the inference time. Last but not least, SEFraud has been deployed and offers explainable fraud detection service for the largest bank in China, Industrial and Commercial Bank of China Limited (ICBC). Results collected from the production environment of ICBC show that SEFraud can provide accurate detection results and comprehensive explanations that align with the expert business understanding, confirming its efficiency and applicability in large-scale online services.

Jiawei Zhou, Xiaoguang Li, Lifeng Shang, Lan Luo, Ke Zhan, Enrui Hu, Xinyu Zhang, Hao Jiang, Zhao Cao, Fan Yu, Xin Jiang, Qun Liu, Lei Chen. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Nowadays, multi-hop question answer (QA) problem is challenging and not well solved in the QA community. The dominant bottleneck of the multi-hop QA problem is the need for a reasoning retriever to fetch a document path from an open-domain corpus (e.g., Wikipedia). A reasoning retriever aims to collect an evidence document from large corpora at one hop retrieval and aggregate the evidence for subsequent hop retrieval, which yields a document path after multi-hop retrieval. There exist two challenges, (1) to fetch the evidence document in an efficient and explainable way at one hop retrieval and (2) to update the question information by aggregating the evidence from the retrieved document after each hop retrieval. To address these two challenges, we propose a triple-fact-based retrieval model to effectively retrieve a related document path in an explainable way for each question. We extract a structured representation from the unstructured document and utilize the knowledge of pre-trained language model (PLM) to do the semantic-level matching between the question and document. We evaluate the proposed Triple-fact Retriever model on the recently proposed open-domain multi-hop QA dataset, HotpotQA, and a cross-document multi-step Reading Comprehension dataset, Wikihop. The results11The source code is available on our website: https://github.com/Rebaccamin/triple_retriever. demonstrate that the Triple-fact retriever outperforms the existing baseline retrieval works.

In recent years, deep learning technology has been used to solve partial differential equations (PDEs), among which the physics-informed neural networks (PINNs)method emerges to be a promising method for solving both forward and inverse PDE problems. PDEs with a point source that is expressed as a Dirac delta function in the governing equations are mathematical models of many physical processes. However, they cannot be solved directly by conventional PINNs method due to the singularity brought by the Dirac delta function. In this paper, we propose a universal solution to tackle this problem by proposing three novel techniques. Firstly the Dirac delta function is modeled as a continuous probability density function to eliminate the singularity at the point source; secondly a lower bound constrained uncertainty weighting algorithm is proposed to balance the physics-informed loss terms of point source area and the remaining areas; and thirdly a multi-scale deep neural network with periodic activation function is used to improve the accuracy and convergence speed. We evaluate the proposed method with three representative PDEs, and the experimental results show that our method outperforms existing deep learning based methods with respect to the accuracy, the efficiency and the versatility.

Many important problems in science and engineering require solving the so-called parametric partial differential equations (PDEs), i.e., PDEs with different physical parameters, boundary conditions, shapes of computation domains, etc. Recently, building learning-based numerical solvers for parametric PDEs has become an emerging new field. One category of methods such as the Deep Galerkin Method (DGM) and Physics-Informed Neural Networks (PINNs) aim to approximate the solution of the PDEs. They are typically unsupervised and mesh-free, but require going through the time-consuming network training process from scratch for each set of parameters of the PDE. Another category of methods such as Fourier Neural Operator (FNO) and Deep Operator Network (DeepONet) try to approximate the solution mapping directly. Being fast with only one forward inference for each PDE parameter without retraining, they often require a large corpus of paired input-output observations drawn from numerical simulations, and most of them need a predefined mesh as well. In this paper, we propose Meta-Auto-Decoder (MAD), a mesh-free and unsupervised deep learning method that enables the pre-trained model to be quickly adapted to equation instances by implicitly encoding (possibly heterogenous) PDE parameters as latent vectors. The proposed method MAD can be interpreted by manifold learning in infinite-dimensional spaces, granting it a geometric insight. Extensive numerical experiments show that the MAD method exhibits faster convergence speed without losing accuracy than other deep learning-based methods.

Graph neural networks (GNNs) have shown to significantly improve graph analytics. Existing systems for GNN training are primarily designed for homogeneous graphs. In industry, however, most graphs are actually heterogeneous in nature (i.e., having multiple types of nodes and edges). Existing systems train a heterogeneous GNN (HetGNN) as a composition of homogeneous GNN (HomoGNN) and thus suffer from critical limitations such as lack of memory optimization and limited operator parallelism. To address these limitations, we propose HGL - a heterogeneity-aware system for GNN training. At the core of HGL is an intermediate representation, called HIR, which provides a holistic representation for GNNs and enables cross-relation optimization in HetGNN training. We devise tailored optimizations on HIR, including graph stitching, operator fusion and operator bundling. Compared with DGL and PyG, HGL achieves a speedup from 7 to 22 times for training HetGNNs.

The configuration of 512 window size prevents transformers from being directly applicable to document ranking that requires larger context. Hence, recent works propose to estimate document relevance with fine-grained passage-level relevance signals. A limitation of such models, however, is that scoring each passage independently falls short in modeling inter-passage interactions and leads to unsatisfactory results. In this paper, we propose a Multiview inter-passage Interaction based Ranking model (MIR), to combine intra-passage interactions and inter-passage interactions in a complementary manner. The former captures local semantic relations inside each passage, whereas the latter draws global dependencies between different passages. Moreover, we represent inter-passage relationships via multi-view attention patterns, allowing information propagation at token, sentence, and passage-level. The representations at different levels of granularity, being aware of global context, are then aggregated into a document-level representation for ranking. Experimental results on two benchmarks show that modeling inter-passage interactions brings substantial improvements over existing passage-level methods.

Effective parallelization strategies are crucial for the performance of distributed deep neural network (DNN) training. Recently, several methods have been proposed to search parallelization strategies but they all optimize a single objective (e.g., execution time, memory consumption) and produce only one strategy. We proposeFrontier Tracking(FT), an efficient algorithm that findsa set of Pareto-optimal parallelization strategiesto explore the best trade-off among different objectives. FT can minimize the memory consumption when the number of devices is limited and fully utilize additional resources to reduce the execution time. Based onFT, we develop a user-friendly system, calledTensorOpt, which allows users to run their distributed DNN training jobs without caring the details about searching and coding parallelization strategies. Experimental results show that TensorOpt is more flexible in adapting to resource availability compared with existing frameworks.

We study how to support elasticity, that is, the ability to dynamically adjust the parallelism (i.e., the number of GPUs), for deep neural network (DNN) training in a GPU cluster. Elasticity can benefit multi-tenant GPU cluster management in many ways, for example, achieving various scheduling objectives (e.g., job throughput, job completion time, GPU efficiency) according to cluster load variations, utilizing transient idle resources, and supporting performance profiling, job migration, and straggler mitigation. We propose EDL, which enables elastic deep learning with a simple API and can be easily integrated with existing deep learning frameworks such as TensorFlow and PyTorch. EDL also incorporates techniques that are necessary to reduce the overhead of parallelism adjustments, such as stop-free scaling and dynamic data pipeline. We demonstrate with experiments that EDL can indeed bring significant benefits to the above-listed applications in GPU cluster management.

It is well-known that second-order optimizer can accelerate the training of deep neural networks, however, the huge computation cost of second-order optimization makes it impractical to apply in real practice. In order to reduce the cost, many methods have been proposed to approximate a second-order matrix. Inspired by KFAC, we propose a novel Trace-based Hardware-driven layer-ORiented Natural Gradient Descent Computation method, called THOR, to make the second-order optimization applicable in the real application models. Specifically, we gradually increase the update interval and use the matrix trace to determine which blocks of Fisher Information Matrix (FIM) need to be updated. Moreover, by resorting the power of hardware, we have designed a Hardware-driven approximation method for computing FIM to achieve better performance. To demonstrate the effectiveness of THOR, we have conducted extensive experiments. The results show that training ResNet-50 on ImageNet with THOR only takes 66.7 minutes to achieve a top-1 accuracy of 75.9 % under an 8 Ascend 910 environment with MindSpore, a new deep learning computing framework. Moreover, with more computational resources, THOR can only takes 2.7 minutes to 75.9 % with 256 Ascend 910.

Second-order optimization methods have the ability to accelerate convergence by modifying the gradient through the curvature matrix. There have been many attempts to use second-order optimization methods for training deep neural networks. In this work, inspired by diagonal approximations and factored approximations such as Kronecker-factored Approximate Curvature (KFAC), we propose a new approximation to the Fisher information matrix (FIM) called Trace-restricted Kronecker-factored Approximate Curvature (TKFAC), which can hold the certain trace relationship between the exact and the approximate FIM. In TKFAC, we decompose each block of the approximate FIM as a Kronecker product of two smaller matrices and scaled by a coefficient related to trace. We theoretically analyze TKFAC's approximation error and give an upper bound of it. We also propose a new damping technique for TKFAC on convolutional neural networks to maintain the superiority of second-order optimization methods during training. Experiments show that our method has better performance compared with several state-of-the-art algorithms on some deep network architectures.

The bottleneck of computation burden limits the widespread use of the 2nd order optimization algorithms for training deep neural networks. In this paper, we present a computationally efficient approximation for natural gradient descent, named Swift Kronecker-Factored Approximate Curvature (SKFAC), which combines Kronecker factorization and a fast low-rank matrix inversion technique. Our research aims at both fully connected and convolutional layers. For the fully connected layers, by utilizing the low-rank property of Kronecker factors of Fisher information matrix, our method only requires inverting a small matrix to approximate the curvature with desirable accuracy. For convolutional layers, we propose a way with two strategies to save computational efforts without affecting the empirical performance by reducing across the spatial dimension or receptive fields of feature maps. Specifically, we propose two effective dimension reduction methods for this purpose: Spatial Subsampling and Reduce Sum. Experimental results of training several deep neural networks on Cifar-10 and ImageNet-1k datasets demonstrate that SKFAC can capture the main curvature and yield comparative performance to K-FAC. The proposed method bridges the wall-clock time gap between the 1st and 2nd order algorithms.

Graph neural networks (GNNs) have gained increasing popularity in many areas such as e-commerce, social networks and bio-informatics. Distributed GNN training is essential for handling large graphs and reducing the execution time. However, for distributed GNN training, a peer-to-peer communication strategy suffers from high communication overheads. Also, different GPUs require different remote vertex embeddings, which leads to an irregular communication pattern and renders existing communication planning solutions unsuitable. We propose the distributed graph communication library (DGCL) for efficient GNN training on multiple GPUs. At the heart of DGCL is a communication planning algorithm tailored for GNN training, which jointly considers fully utilizing fast links, fusing communication, avoiding contention and balancing loads on different links. DGCL can be easily adopted to extend existing single-GPU GNN systems to distributed training. We conducted extensive experiments on different datasets and network configurations to compare DGCL with alternative communication schemes. In our experiments, DGCL reduces the communication time of the peer-to-peer communication by 77.5% on average and the training time for an epoch by up to 47%.

Graph neural networks (GNNs) have achieved breakthrough performance in graph analytics such as node classification, link prediction and graph clustering. Many GNN training frameworks have been developed, but they are usually designed as a set of manually written, GNN-specific operators plugged into existing deep learning systems, which incurs high memory consumption, poor data locality, and large semantic gap between algorithm design and implementation. This paper proposes the Seastar system, which presents a vertex-centric programming model for GNN training on GPU and provides idiomatic python constructs to enable easy development of novel homogeneous and heterogeneous GNN models. We also propose novel optimizations to produce highly efficient fused GPU kernels for forward and backward passes in GNN training. Compared with the state-of-the art GNN systems, DGL and PyG, Seastar achieves better usability, up to 2 and 8 times less memory consumption, and 14 and 3 times faster execution, respectively.

In this paper, we proposed EPP-MVSNet, a novel deep learning network for 3D reconstruction from multi-view stereo (MVS). EPP-MVSNet can accurately aggregate features at high resolution to a limited cost volume with an optimal depth range, thus, leads to effective and efficient 3D construction. Distinct from existing works which measure feature cost at discrete positions which affects the 3D reconstruction accuracy, EPP-MVSNet introduces an epipolar-assembling-based kernel that operates on adaptive intervals along epipolar lines for making full use of the image resolution. Further, we introduce an entropy-based refining strategy where the cost volume describes the space geometry with the little redundancy. Moreover, we design a light-weighted network with Pseudo-3D convolutions integrated to achieve high accuracy and efficiency. We have conducted extensive experiments on challenging datasets Tanks & Temples(TNT), ETH3D and DTU. As a result, we achieve promising results on all datasets and the highest F-Score on the online TNT intermediate benchmark. Code is available at https://gitee.com/mindspore/mindspore/tree/master/model_zoo/research/cv/eppmvsnet.

Currently, the most popular method for open-domain Question Answering (QA) adopts "Retriever and Reader" pipeline, where the retriever extracts a list of candidate documents from a large set of documents followed by a ranker to rank the most relevant documents and the reader extracts answer from the candidates. Existing studies take the greedy strategy in the sense that they only use samples for ranking at the current hop, and ignore the global information across the whole documents. In this paper, we propose a purely rank-based framework Thinking Path Re-Ranker (TPRR), which is comprised of Thinking Path Ranker (TPR) for generating document sequences called "a path" and External Path Reranker (EPR) for selecting the best path from candidate paths generated by TPR. Specifically, TPR leverages the scores of a dense model and conditional probabilities to score the full paths. Moreover, to further enhance the performance of the dense ranker in the iterative training, we propose a "thinking" negatives selection method that the top-K candidates treated as negatives in the current hop are adjusted dynamically through supervised signals. After achieving multiple supporting paths through TPR, the EPR component which integrates several fine-grained training tasks for QA is used to select the best path for answer extraction. We have tested our proposed solution on the multi-hop dataset "HotpotQA" with a full wiki set ting, and the results show that TPRR significantly outperforms the existing state-of-the-art models. Moreover, our method has won the first place in the HotpotQA official leaderboard since Feb 1, 2021 under the Fullwiki setting. Code is available at https://gitee.com/mindspore/mindspore/ tree/master/model_zoo/research/nlp/tprr.

Graph neural networks (GNNs) have achieved remarkable performance in many graph analytics tasks such as node classification, link prediction and graph clustering. Existing GNN systems (e.g., PyG and DGL) adopt a tensor-centric programming model and train GNNs with manually written operators. Such design results in poor usability due to the large semantic gap between the API and the GNN models, and suffers from inferior efficiency because of high memory consumption and massive data movement. We demonstrateSeastar, a novel GNN training framework that adopts avertex-centric programming paradigm and supportsautomatic kernel generation, to simplify model development and improve training efficiency. We will (i) show how to express GNN models succinctly using a visual "drag-and-drop'' interface or Seastar's vertex-centric python API; (ii) demonstrate the performance advantage of Seastar over existing GNN systems in convergence speed, training throughput and memory consumption; and (iii) illustrate how Seastar's optimizations (e.g., operator fusion and constant folding) improve training efficiency by profiling the run-time performance.

For the action localization task, pre-defined action anchors are the cornerstone of mainstream techniques. State-of-the-art models mostly rely on a dense segmenting scheme, where anchors are sampled uniformly over the temporal domain with a predefined set of scales. However, it is not sufficient because action duration varies greatly. Therefore, it is necessary for the anchors or proposals to have a variable receptive field. In this paper, we propose a method called YOLA (You Ought to Look Around) which includes three parts: 1) a robust backbone SPN-I3D for extracting spatio-temporal features. In this part, we employ a stronger backbone I3D with SPN (Segment Pyramid Network) instead of C3D to obtain multi-scale features; 2) a simple but useful feature fusion module named LFE (Local Feature Extraction). Compared with the fully connected layer and global average pooling, our LFE model is more advantageous for network to fit and fuse features. 3) a new feature segment aligning method called TPGC (Two Pathway Graph Convolution), which allows one proposal to leverage semantic features of adjacent proposals to update its content and make sure the proposals have a variable receptive field. YOLA add only a small overhead to the baseline network, and is easy to train in an end-to-end manner, running at a speed of 1097 fps. YOLA achieves a mAP of 58.3%, outperforming all existing models including both RGB-based and two stream on THUMOS'14, and achieves competitive results on ActivityNet 1.3.