Fan Yu

Joey Wang, Yingcan Wei, Minseok Lee et al.

In this talk, we introduce Merlin HugeCTR. Merlin HugeCTR is an open source, GPU-accelerated integration framework for click-through rate estimation. It optimizes both training and inference, whilst enabling model training at scale with model-parallel embeddings and data-parallel neural networks. In particular, Merlin HugeCTR combines a high-performance GPU embedding cache with an hierarchical storage architecture, to realize low-latency retrieval of embeddings for online model inference tasks. In the MLPerf v1.0 DLRM model training benchmark, Merlin HugeCTR achieves a speedup of up to 24.6x on a single DGX A100 (8x A100) over PyTorch on 4x4-socket CPU nodes (4x4x28 cores). Merlin HugeCTR can also take advantage of multi-node environments to accelerate training even further. Since late 2021, Merlin HugeCTR additionally features a hierarchical parameter server (HPS) and supports deployment via the NVIDIA Triton server framework, to leverage the computational capabilities of GPUs for high-speed recommendation model inference. Using this HPS, Merlin HugeCTR users can achieve a 5~62x speedup (batch size dependent) for popular recommendation models over CPU baseline implementations, and dramatically reduce their end-to-end inference latency.

Sirui Liu, Jun Zhang, Haotian Chu et al.

Proteins are essential component of human life and their structures are important for function and mechanism analysis. Recent work has shown the potential of AI-driven methods for protein structure prediction. However, the development of new models is restricted by the lack of dataset and benchmark training procedure. To the best of our knowledge, the existing open source datasets are far less to satisfy the needs of modern protein sequence-structure related research. To solve this problem, we present the first million-level protein structure prediction dataset with high coverage and diversity, named as PSP. This dataset consists of 570k true structure sequences (10TB) and 745k complementary distillation sequences (15TB). We provide in addition the benchmark training procedure for SOTA protein structure prediction model on this dataset. We validate the utility of this dataset for training by participating CAMEO contest in which our model won the first place. We hope our PSP dataset together with the training benchmark can enable a broader community of AI/biology researchers for AI-driven protein related research.

Jiawei Zhou, Xiaoguang Li, Lifeng Shang et al.

To alleviate the data scarcity problem in training question answering systems, recent works propose additional intermediate pre-training for dense passage retrieval (DPR). However, there still remains a large discrepancy between the provided upstream signals and the downstream question-passage relevance, which leads to less improvement. To bridge this gap, we propose the HyperLink-induced Pre-training (HLP), a method to pre-train the dense retriever with the text relevance induced by hyperlink-based topology within Web documents. We demonstrate that the hyperlink-based structures of dual-link and co-mention can provide effective relevance signals for large-scale pre-training that better facilitate downstream passage retrieval. We investigate the effectiveness of our approach across a wide range of open-domain QA datasets under zero-shot, few-shot, multi-hop, and out-of-domain scenarios. The experiments show our HLP outperforms the BM25 by up to 7 points as well as other pre-training methods by more than 10 points in terms of top-20 retrieval accuracy under the zero-shot scenario. Furthermore, HLP significantly outperforms other pre-training methods under the other scenarios.

Haidong Rong, Jiashu Yao, Matthias Langer et al.

Traditional GPU hash tables preserve every inserted key -- a dictionary assumption that wastes scarce High Bandwidth Memory (HBM) when embedding tables routinely exceed single-GPU capacity. We challenge this assumption with cache semantics, where policy-driven eviction is a first-class operation. We introduce HierarchicalKV (HKV), the first general-purpose GPU hash table library whose normal full-capacity operating contract is cache-semantic: each full-bucket upsert (update-or-insert) is resolved in place by eviction or admission rejection rather than by rehashing or capacity-induced failure. HKV co-designs four core mechanisms -- cache-line-aligned buckets, in-line score-driven upsert, score-based dynamic dual-bucket selection, and triple-group concurrency -- and uses tiered key-value separation as a scaling enabler beyond HBM. On an NVIDIA H100 NVL GPU, HKV achieves up to 3.9 billion key-value pairs per second (B-KV/s) find throughput, stable across load factors 0.50-1.00 (<5% variation), and delivers 1.4x higher find throughput than WarpCore (the strongest dictionary-semantic GPU baseline at lambda=0.50) and up to 2.6-9.4x over indirection-based GPU baselines. Since its open-source release in October 2022, HKV has been integrated into multiple open-source recommendation frameworks.

Yingcan Wei, Matthias Langer, Fan Yu et al.

Recommendation systems are of crucial importance for a variety of modern apps and web services, such as news feeds, social networks, e-commerce, search, etc. To achieve peak prediction accuracy, modern recommendation models combine deep learning with terabyte-scale embedding tables to obtain a fine-grained representation of the underlying data. Traditional inference serving architectures require deploying the whole model to standalone servers, which is infeasible at such massive scale. In this paper, we provide insights into the intriguing and challenging inference domain of online recommendation systems. We propose the HugeCTR Hierarchical Parameter Server (HPS), an industry-leading distributed recommendation inference framework, that combines a high-performance GPU embedding cache with an hierarchical storage architecture, to realize low-latency retrieval of embeddings for online model inference tasks. Among other things, HPS features (1) a redundant hierarchical storage system, (2) a novel high-bandwidth cache to accelerate parallel embedding lookup on NVIDIA GPUs, (3) online training support and (4) light-weight APIs for easy integration into existing large-scale recommendation workflows. To demonstrate its capabilities, we conduct extensive studies using both synthetically engineered and public datasets. We show that our HPS can dramatically reduce end-to-end inference latency, achieving 5~62x speedup (depending on the batch size) over CPU baseline implementations for popular recommendation models. Through multi-GPU concurrent deployment, the HPS can also greatly increase the inference QPS.

Jun Zhang, Sirui Liu, Mengyun Chen et al.

Data-driven predictive methods which can efficiently and accurately transform protein sequences into biologically active structures are highly valuable for scientific research and medical development. Determining accurate folding landscape using co-evolutionary information is fundamental to the success of modern protein structure prediction methods. As the state of the art, AlphaFold2 has dramatically raised the accuracy without performing explicit co-evolutionary analysis. Nevertheless, its performance still shows strong dependence on available sequence homologs. Based on the interrogation on the cause of such dependence, we presented EvoGen, a meta generative model, to remedy the underperformance of AlphaFold2 for poor MSA targets. By prompting the model with calibrated or virtually generated homologue sequences, EvoGen helps AlphaFold2 fold accurately in low-data regime and even achieve encouraging performance with single-sequence predictions. Being able to make accurate predictions with few-shot MSA not only generalizes AlphaFold2 better for orphan sequences, but also democratizes its use for high-throughput applications. Besides, EvoGen combined with AlphaFold2 yields a probabilistic structure generation method which could explore alternative conformations of protein sequences, and the task-aware differentiable algorithm for sequence generation will benefit other related tasks including protein design.

Amos Goldman, Nimrod Boker, Maayan Sheraizin et al.

Mixture-of-Experts (MoE) architectures have become essential for scaling large language models, driving the development of specialized device-initiated communication libraries such as DeepEP, Hybrid-EP, and others. These libraries demonstrate the performance benefits of GPU-initiated RDMA for MoE dispatch and combine operations. This paper presents NCCL EP (Expert Parallelism), a ground-up MoE communication library built entirely on NCCL's Device API. NCCL EP provides unified ncclEpDispatch and ncclEpCombine primitives with both C and Python interfaces, supporting Low-Latency (LL) mode for inference decoding and High-Throughput (HT) mode for training and inference prefill. LL targets small batch sizes (1-128 tokens) using direct all-to-all RDMA+NVLink mesh connectivity with double-buffered communication for overlapping dispatch and combine phases. HT targets large batches (4096+ tokens) using hierarchical communication that aggregates tokens within NVLink domains before inter-node RDMA transmission. Both modes leverage Device API for both intra- and inter-node communications, taking advantage of its topology awareness and optimized GPU-initiated implementation. We evaluate NCCL EP on an H100-based cluster across multi-node configurations, demonstrating competitive LL kernel performance and presenting end-to-end results with vLLM integration. By building MoE communication natively within NCCL, NCCL EP provides a supported path for expert parallelism on current and emerging NVIDIA platforms.

Guanrou Yang, Chen Yang, Qian Chen et al.

Human speech goes beyond the mere transfer of information; it is a profound exchange of emotions and a connection between individuals. While Text-to-Speech (TTS) models have made huge progress, they still face challenges in controlling the emotional expression in the generated speech. In this work, we propose EmoVoice, a novel emotion-controllable TTS model that exploits large language models (LLMs) to enable fine-grained freestyle natural language emotion control, and a phoneme boost variant design that makes the model output phoneme tokens and audio tokens in parallel to enhance content consistency, inspired by chain-of-thought (CoT) and chain-of-modality (CoM) techniques. Besides, we introduce EmoVoice-DB, a high-quality 40-hour English emotion dataset featuring expressive speech and fine-grained emotion labels with natural language descriptions. EmoVoice achieves state-of-the-art performance on the English EmoVoice-DB test set using only synthetic training data, and on the Chinese Secap test set using our in-house data. We further investigate the reliability of existing emotion evaluation metrics and their alignment with human perceptual preferences, and explore using SOTA multimodal LLMs GPT-4o-audio and Gemini to assess emotional speech. Dataset, code, checkpoints, and demo samples are available at https://github.com/yanghaha0908/EmoVoice.

Zhihao Du, Yuxuan Wang, Qian Chen et al.

In our previous work, we introduced CosyVoice, a multilingual speech synthesis model based on supervised discrete speech tokens. By employing progressive semantic decoding with two popular generative models, language models (LMs) and Flow Matching, CosyVoice demonstrated high prosody naturalness, content consistency, and speaker similarity in speech in-context learning. Recently, significant progress has been made in multi-modal large language models (LLMs), where the response latency and real-time factor of speech synthesis play a crucial role in the interactive experience. Therefore, in this report, we present an improved streaming speech synthesis model, CosyVoice 2, which incorporates comprehensive and systematic optimizations. Specifically, we introduce finite-scalar quantization to improve the codebook utilization of speech tokens. For the text-speech LM, we streamline the model architecture to allow direct use of a pre-trained LLM as the backbone. In addition, we develop a chunk-aware causal flow matching model to support various synthesis scenarios, enabling both streaming and non-streaming synthesis within a single model. By training on a large-scale multilingual dataset, CosyVoice 2 achieves human-parity naturalness, minimal response latency, and virtually lossless synthesis quality in the streaming mode. We invite readers to listen to the demos at https://funaudiollm.github.io/cosyvoice2.

Xiang Huang, Zhanhong Ye, Hongsheng Liu et al.

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. The project page with code is available: https://gitee.com/mindspore/mindscience/tree/master/MindElec/.

Fan Yu, Haoneng Luo, Pengcheng Guo et al.

Continuous integrate-and-fire (CIF) based models, which use a soft and monotonic alignment mechanism, have been well applied in non-autoregressive (NAR) speech recognition with competitive performance compared with other NAR methods. However, such an alignment learning strategy may suffer from an erroneous acoustic boundary estimation, severely hindering the convergence speed as well as the system performance. In this paper, we propose a boundary and context aware training approach for CIF based NAR models. Firstly, the connectionist temporal classification (CTC) spike information is utilized to guide the learning of acoustic boundaries in the CIF. Besides, an additional contextual decoder is introduced behind the CIF decoder, aiming to capture the linguistic dependencies within a sentence. Finally, we adopt a recently proposed Conformer architecture to improve the capacity of acoustic modeling. Experiments on the open-source Mandarin AISHELL-1 corpus show that the proposed method achieves a comparable character error rates (CERs) of 4.9% with only 1/24 latency compared with a state-of-the-art autoregressive (AR) Conformer model. Futhermore, when evaluating on an internal 7500 hours Mandarin corpus, our model still outperforms other NAR methods and even reaches the AR Conformer model on a challenging real-world noisy test set.

Zhuoyuan Yao, Di Wu, Xiong Wang et al.

In this paper, we propose an open source, production first, and production ready speech recognition toolkit called WeNet in which a new two-pass approach is implemented to unify streaming and non-streaming end-to-end (E2E) speech recognition in a single model. The main motivation of WeNet is to close the gap between the research and the production of E2E speechrecognition models. WeNet provides an efficient way to ship ASR applications in several real-world scenarios, which is the main difference and advantage to other open source E2E speech recognition toolkits. In our toolkit, a new two-pass method is implemented. Our method propose a dynamic chunk-based attention strategy of the the transformer layers to allow arbitrary right context length modifies in hybrid CTC/attention architecture. The inference latency could be easily controlled by only changing the chunk size. The CTC hypotheses are then rescored by the attention decoder to get the final result. Our experiments on the AISHELL-1 dataset using WeNet show that, our model achieves 5.03\% relative character error rate (CER) reduction in non-streaming ASR compared to a standard non-streaming transformer. After model quantification, our model perform reasonable RTF and latency.

Ziyang Ma, Guanrou Yang, Yifan Yang et al.

In this paper, we focus on solving one of the most important tasks in the field of speech processing, i.e., automatic speech recognition (ASR), with speech foundation encoders and large language models (LLM). Recent works have complex designs such as compressing the output temporally for the speech encoder, tackling modal alignment for the projector, and utilizing parameter-efficient fine-tuning for the LLM. We found that delicate designs are not necessary, while an embarrassingly simple composition of off-the-shelf speech encoder, LLM, and the only trainable linear projector is competent for the ASR task. To be more specific, we benchmark and explore various combinations of LLMs and speech encoders, leading to the optimal LLM-based ASR system, which we call SLAM-ASR. The proposed SLAM-ASR provides a clean setup and little task-specific design, where only the linear projector is trained. To the best of our knowledge, SLAM-ASR achieves the best performance on the Librispeech benchmark among LLM-based ASR models and even outperforms the latest LLM-based audio-universal model trained on massive pair data. Finally, we explore the capability emergence of LLM-based ASR in the process of modal alignment. We hope that our study can facilitate the research on extending LLM with cross-modality capacity and shed light on the LLM-based ASR community.

Qian Chen, Yafeng Chen, Yanni Chen et al.

Recent advancements in large language models (LLMs) and multimodal speech-text models have laid the groundwork for seamless voice interactions, enabling real-time, natural, and human-like conversations. Previous models for voice interactions are categorized as native and aligned. Native models integrate speech and text processing in one framework but struggle with issues like differing sequence lengths and insufficient pre-training. Aligned models maintain text LLM capabilities but are often limited by small datasets and a narrow focus on speech tasks. In this work, we introduce MinMo, a Multimodal Large Language Model with approximately 8B parameters for seamless voice interaction. We address the main limitations of prior aligned multimodal models. We train MinMo through multiple stages of speech-to-text alignment, text-to-speech alignment, speech-to-speech alignment, and duplex interaction alignment, on 1.4 million hours of diverse speech data and a broad range of speech tasks. After the multi-stage training, MinMo achieves state-of-the-art performance across various benchmarks for voice comprehension and generation while maintaining the capabilities of text LLMs, and also facilitates full-duplex conversation, that is, simultaneous two-way communication between the user and the system. Moreover, we propose a novel and simple voice decoder that outperforms prior models in voice generation. The enhanced instruction-following capabilities of MinMo supports controlling speech generation based on user instructions, with various nuances including emotions, dialects, and speaking rates, and mimicking specific voices. For MinMo, the speech-to-text latency is approximately 100ms, full-duplex latency is approximately 600ms in theory and 800ms in practice. The MinMo project web page is https://funaudiollm.github.io/minmo, and the code and models will be released soon.

Ke Wang, Yanfei Cao, Xiangzhi Tao et al.

With the rapid development of computer vision and deep learning, significant advancements have been made in 3D vision, partic- ularly in autonomous driving, robotic perception, and augmented reality. 3D point cloud data, as a crucial representation of 3D information, has gained widespread attention. However, the vast scale and complexity of point cloud data present significant chal- lenges for loading and processing and traditional algorithms struggle to handle large-scale datasets.The diversity of storage formats for point cloud datasets (e.g., PLY, XYZ, BIN) adds complexity to data handling and results in inefficiencies in data preparation. Al- though binary formats like BIN and NPY have been used to speed up data access, they still do not fully address the time-consuming data loading and processing phase. To overcome these challenges, we propose the .PcRecord format, a unified data storage solution designed to reduce the storage occupation and accelerate the processing of point cloud data. We also introduce a high-performance data processing pipeline equipped with multiple modules. By leveraging a multi-stage parallel pipeline architecture, our system optimizes the use of computational resources, significantly improving processing speed and efficiency. This paper details the im- plementation of this system and demonstrates its effectiveness in addressing the challenges of handling large-scale point cloud datasets.On average, our system achieves performance improvements of 6.61x (ModelNet40), 2.69x (S3DIS), 2.23x (ShapeNet), 3.09x (Kitti), 8.07x (SUN RGB-D), and 5.67x (ScanNet) with GPU and 6.9x, 1.88x, 1.29x, 2.28x, 25.4x, and 19.3x with Ascend.

Zhihao Du, Changfeng Gao, Yuxuan Wang et al.

In our prior works, we introduced a scalable streaming speech synthesis model, CosyVoice 2, which integrates a large language model (LLM) and a chunk-aware flow matching (FM) model, and achieves low-latency bi-streaming speech synthesis and human-parity quality. Despite these advancements, CosyVoice 2 exhibits limitations in language coverage, domain diversity, data volume, text formats, and post-training techniques. In this paper, we present CosyVoice 3, an improved model designed for zero-shot multilingual speech synthesis in the wild, surpassing its predecessor in content consistency, speaker similarity, and prosody naturalness. Key features of CosyVoice 3 include: 1) A novel speech tokenizer to improve prosody naturalness, developed via supervised multi-task training, including automatic speech recognition, speech emotion recognition, language identification, audio event detection, and speaker analysis. 2) A new differentiable reward model for post-training applicable not only to CosyVoice 3 but also to other LLM-based speech synthesis models. 3) Dataset Size Scaling: Training data is expanded from ten thousand hours to one million hours, encompassing 9 languages and 18 Chinese dialects across various domains and text formats. 4) Model Size Scaling: Model parameters are increased from 0.5 billion to 1.5 billion, resulting in enhanced performance on our multilingual benchmark due to the larger model capacity. These advancements contribute significantly to the progress of speech synthesis in the wild. We encourage readers to listen to the demo at https://funaudiollm.github.io/cosyvoice3.

Guanrou Yang, Fan Yu, Ziyang Ma et al.

While automatic speech recognition (ASR) systems have achieved remarkable performance with large-scale datasets, their efficacy remains inadequate in low-resource settings, encompassing dialects, accents, minority languages, and long-tail hotwords, domains with significant practical relevance. With the advent of versatile and powerful text-to-speech (TTS) models, capable of generating speech with human-level naturalness, expressiveness, and diverse speaker profiles, leveraging TTS for ASR data augmentation provides a cost-effective and practical approach to enhancing ASR performance. Comprehensive experiments on an unprecedentedly rich variety of low-resource datasets demonstrate consistent and substantial performance improvements, proving that the proposed method of enhancing low-resource ASR through a versatile TTS model is highly effective and has broad application prospects. Furthermore, we delve deeper into key characteristics of synthesized speech data that contribute to ASR improvement, examining factors such as text diversity, speaker diversity, and the volume of synthesized data, with text diversity being studied for the first time in this work. We hope our findings provide helpful guidance and reference for the practical application of TTS-based data augmentation and push the advancement of low-resource ASR one step further.

Yiming Sun, Fan Yu, Shaoxiang Chen et al.

Visual object tracking aims to locate a targeted object in a video sequence based on an initial bounding box. Recently, Vision-Language~(VL) trackers have proposed to utilize additional natural language descriptions to enhance versatility in various applications. However, VL trackers are still inferior to State-of-The-Art (SoTA) visual trackers in terms of tracking performance. We found that this inferiority primarily results from their heavy reliance on manual textual annotations, which include the frequent provision of ambiguous language descriptions. In this paper, we propose ChatTracker to leverage the wealth of world knowledge in the Multimodal Large Language Model (MLLM) to generate high-quality language descriptions and enhance tracking performance. To this end, we propose a novel reflection-based prompt optimization module to iteratively refine the ambiguous and inaccurate descriptions of the target with tracking feedback. To further utilize semantic information produced by MLLM, a simple yet effective VL tracking framework is proposed and can be easily integrated as a plug-and-play module to boost the performance of both VL and visual trackers. Experimental results show that our proposed ChatTracker achieves a performance comparable to existing methods.

Zhiyuan Fang, Yuegui Huang, Zicong Hong et al.

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. Therefore, we propose Klotski, an efficient MoE inference engine that significantly reduces pipeline bubbles through a novel expert-aware multi-batch pipeline paradigm. The proposed paradigm uses batch processing to extend the computation time of the current layer to overlap with the loading time of the next layer. Although this idea has been effectively applied to dense models, more batches may activate more experts in the MoE, leading to longer loading times and more bubbles. Thus, unlike traditional approaches, we balance computation and I/O time and minimise bubbles by orchestrating their inference orders based on their heterogeneous computation and I/O requirements and activation patterns under different batch numbers. Moreover, to adapt to different hardware environments and models, we design a constraint-sensitive I/O-compute planner and a correlation-aware expert prefetcher for a schedule that minimises pipeline bubbles. Experimental results demonstrate that Klotski achieves a superior throughput-latency trade-off compared to state-of-the-art techniques, with throughput improvements of up to 85.12x.

Mengtao Yan, Qi Wang, Haining Wang et al.

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.

Jian Li, Wan Han, Ning Lin et al.

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.

Zhiyuan Fang, Zicong Hong, Yuegui Huang et al.

Large Language Models (LLMs) have demonstrated impressive performance across various tasks, and their application in edge scenarios has attracted significant attention. However, sparse-activated Mixture-of-Experts (MoE) models, which are well suited for edge scenarios, have received relatively little attention due to their high memory demands. Offload-based methods have been proposed to address this challenge, but they face difficulties with expert prediction. Inaccurate expert predictions can result in prolonged inference delays. 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 without additional GPU overhead. 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 IO efficiency. Experimental results show that, compared to Load on Demand and Expert Activation Path-based method, Fate achieves up to 4.5x and 1.9x speedups in prefill speed and up to 4.1x and 2.2x speedups in decoding speed, respectively, while maintaining inference quality. Moreover, Fate's performance improvements are scalable across different memory budgets.

Bo Cao, Fan Yu, Mengmeng Feng et al.

Multimodal learning has attracted much attention in recent years due to its ability to effectively utilize data features from a variety of different modalities. Diagnosing the vulnerability of atherosclerotic plaques directly from carotid 3D MRI images is relatively challenging for both radiologists and conventional 3D vision networks. In clinical practice, radiologists assess patient conditions using a multimodal approach that incorporates various imaging modalities and domain-specific expertise, paving the way for the creation of multimodal diagnostic networks. In this paper, we have developed an effective strategy to leverage radiologists' domain knowledge to automate the diagnosis of carotid plaque vulnerability through Variation inference and Multimodal knowledge Distillation (VMD). This method excels in harnessing cross-modality prior knowledge from limited image annotations and radiology reports within training data, thereby enhancing the diagnostic network's accuracy for unannotated 3D MRI images. We conducted in-depth experiments on the dataset collected in-house and verified the effectiveness of the VMD strategy we proposed.

Kaidi Li, Tianmeng Yang, Min Zhou et al.

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. Moreover, 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.

Guanrou Yang, Ziyang Ma, Fan Yu et al.

As more and more information-rich data like video become available, utilizing multi-modal auxiliary information to enhance audio tasks has sparked widespread research interest. The recent surge in research on LLM-based audio models provides fresh perspectives for tackling audio tasks. Given that LLM can flexibly ingest multiple inputs, we propose MaLa-ASR, an LLM-based ASR model that can integrate textual keywords extracted from presentation slides to improve recognition of conference content. MaLa-ASR yields average WERs of 9.4% and 11.7% on the L95 and S95 subsets of the SlideSpeech corpus, representing a significant relative WER drop of 27.9% and 44.7% over the baseline model reported in SlideSpeech. MaLa-ASR underscores LLM's strong performance in speech tasks and the capability to integrate auxiliary information conveniently. By adding keywords to the input prompt, the biased word error rate (B-WER) reduces relatively by 46.0% and 44.2%, establishing a new SOTA on this dataset.

Yuhao Liang, Fan Yu, Yangze Li et al.

The recently proposed serialized output training (SOT) simplifies multi-talker automatic speech recognition (ASR) by generating speaker transcriptions separated by a special token. However, frequent speaker changes can make speaker change prediction difficult. To address this, we propose boundary-aware serialized output training (BA-SOT), which explicitly incorporates boundary knowledge into the decoder via a speaker change detection task and boundary constraint loss. We also introduce a two-stage connectionist temporal classification (CTC) strategy that incorporates token-level SOT CTC to restore temporal context information. Besides typical character error rate (CER), we introduce utterance-dependent character error rate (UD-CER) to further measure the precision of speaker change prediction. Compared to original SOT, BA-SOT reduces CER/UD-CER by 5.1%/14.0%, and leveraging a pre-trained ASR model for BA-SOT model initialization further reduces CER/UD-CER by 8.4%/19.9%.

Fan Yu, Zhihao Du, Shiliang Zhang et al.

In this paper, we conduct a comparative study on speaker-attributed automatic speech recognition (SA-ASR) in the multi-party meeting scenario, a topic with increasing attention in meeting rich transcription. Specifically, three approaches are evaluated in this study. The first approach, FD-SOT, consists of a frame-level diarization model to identify speakers and a multi-talker ASR to recognize utterances. The speaker-attributed transcriptions are obtained by aligning the diarization results and recognized hypotheses. However, such an alignment strategy may suffer from erroneous timestamps due to the modular independence, severely hindering the model performance. Therefore, we propose the second approach, WD-SOT, to address alignment errors by introducing a word-level diarization model, which can get rid of such timestamp alignment dependency. To further mitigate the alignment issues, we propose the third approach, TS-ASR, which trains a target-speaker separation module and an ASR module jointly. By comparing various strategies for each SA-ASR approach, experimental results on a real meeting scenario corpus, AliMeeting, reveal that the WD-SOT approach achieves 10.7% relative reduction on averaged speaker-dependent character error rate (SD-CER), compared with the FD-SOT approach. In addition, the TS-ASR approach also outperforms the FD-SOT approach and brings 16.5% relative average SD-CER reduction.

Fan Yu, Shiliang Zhang, Pengcheng Guo et al.

The ICASSP 2022 Multi-channel Multi-party Meeting Transcription Grand Challenge (M2MeT) focuses on one of the most valuable and the most challenging scenarios of speech technologies. The M2MeT challenge has particularly set up two tracks, speaker diarization (track 1) and multi-speaker automatic speech recognition (ASR) (track 2). Along with the challenge, we released 120 hours of real-recorded Mandarin meeting speech data with manual annotation, including far-field data collected by 8-channel microphone array as well as near-field data collected by each participants' headset microphone. We briefly describe the released dataset, track setups, baselines and summarize the challenge results and major techniques used in the submissions.

Xiang Huang, Hongsheng Liu, Beiji Shi et al.

In recent years, deep learning technology has been used to solve partial differential equations (PDEs), among which the physics-informed neural networks (PINNs) 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. We propose a universal solution to tackle this problem with three novel techniques. Firstly the Dirac delta function is modeled as a continuous probability density function to eliminate the singularity; secondly a lower bound constrained uncertainty weighting algorithm is proposed to balance the PINNs losses between point source area and other areas; and thirdly a multi-scale deep neural network with periodic activation function is used to improve the accuracy and convergence speed of the PINNs method. 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.

Hao Jiang, Ke Zhan, Jianwei Qu et al.

The sparsely-activated models have achieved great success in natural language processing through large-scale parameters and relatively low computational cost, and gradually become a feasible technique for training and implementing extremely large models. Due to the limit of communication cost, activating multiple experts is hardly affordable during training and inference. Therefore, previous work usually activate just one expert at a time to alleviate additional communication cost. Such routing mechanism limits the upper bound of model performance. In this paper, we first investigate a phenomenon that increasing the number of activated experts can boost the model performance with higher sparse ratio. To increase the number of activated experts without an increase in computational cost, we propose SAM (Switch and Mixture) routing, an efficient hierarchical routing mechanism that activates multiple experts in a same device (GPU). Our methods shed light on the training of extremely large sparse models and experiments prove that our models can achieve significant performance gain with great efficiency improvement.

Fan Yu, Shiliang Zhang, Yihui Fu et al.

Recent development of speech processing, such as speech recognition, speaker diarization, etc., has inspired numerous applications of speech technologies. The meeting scenario is one of the most valuable and, at the same time, most challenging scenarios for the deployment of speech technologies. Specifically, two typical tasks, speaker diarization and multi-speaker automatic speech recognition have attracted much attention recently. However, the lack of large public meeting data has been a major obstacle for the advancement of the field. Therefore, we make available the AliMeeting corpus, which consists of 120 hours of recorded Mandarin meeting data, including far-field data collected by 8-channel microphone array as well as near-field data collected by headset microphone. Each meeting session is composed of 2-4 speakers with different speaker overlap ratio, recorded in rooms with different size. Along with the dataset, we launch the ICASSP 2022 Multi-channel Multi-party Meeting Transcription Challenge (M2MeT) with two tracks, namely speaker diarization and multi-speaker ASR, aiming to provide a common testbed for meeting rich transcription and promote reproducible research in this field. In this paper we provide a detailed introduction of the AliMeeting dateset, challenge rules, evaluation methods and baseline systems.

Xian Shi, Fan Yu, Yizhou Lu et al.

The variety of accents has posed a big challenge to speech recognition. The Accented English Speech Recognition Challenge (AESRC2020) is designed for providing a common testbed and promoting accent-related research. Two tracks are set in the challenge -- English accent recognition (track 1) and accented English speech recognition (track 2). A set of 160 hours of accented English speech collected from 8 countries is released with labels as the training set. Another 20 hours of speech without labels is later released as the test set, including two unseen accents from another two countries used to test the model generalization ability in track 2. We also provide baseline systems for the participants. This paper first reviews the released dataset, track setups, baselines and then summarizes the challenge results and major techniques used in the submissions.

Zedong Tang, Fenlong Jiang, Junke Song et al.

Optimizers that further adjust the scale of gradient, such as Adam, Natural Gradient (NG), etc., despite widely concerned and used by the community, are often found poor generalization performance, compared with Stochastic Gradient Descent (SGD). They tend to converge excellently at the beginning of training but are weak at the end. An immediate idea is to complement the strengths of these algorithms with SGD. However, a truncated replacement of optimizer often leads to a crash of the update pattern, and new algorithms often require many iterations to stabilize their search direction. Driven by this idea and to address this problem, we design and present a regularized natural gradient optimization algorithm with look-ahead strategy, named asymptotic natural gradient (ANG). According to the total iteration step, ANG dynamic assembles NG and Euclidean gradient, and updates parameters along the new direction using the intensity of NG. Validation experiments on CIFAR10 and CIFAR100 data sets show that ANG can update smoothly and stably at the second-order speed, and achieve better generalization performance.

Binbin Zhang, Di Wu, Zhuoyuan Yao et al.

In this paper, we present a novel two-pass approach to unify streaming and non-streaming end-to-end (E2E) speech recognition in a single model. Our model adopts the hybrid CTC/attention architecture, in which the conformer layers in the encoder are modified. We propose a dynamic chunk-based attention strategy to allow arbitrary right context length. At inference time, the CTC decoder generates n-best hypotheses in a streaming way. The inference latency could be easily controlled by only changing the chunk size. The CTC hypotheses are then rescored by the attention decoder to get the final result. This efficient rescoring process causes very little sentence-level latency. Our experiments on the open 170-hour AISHELL-1 dataset show that, the proposed method can unify the streaming and non-streaming model simply and efficiently. On the AISHELL-1 test set, our unified model achieves 5.60% relative character error rate (CER) reduction in non-streaming ASR compared to a standard non-streaming transformer. The same model achieves 5.42% CER with 640ms latency in a streaming ASR system.

Kai-Xin Gao, Xiao-Lei Liu, Zheng-Hai Huang et al.

Using second-order optimization methods for training deep neural networks (DNNs) has attracted many researchers. A recently proposed method, Eigenvalue-corrected Kronecker Factorization (EKFAC) (George et al., 2018), proposes an interpretation of viewing natural gradient update as a diagonal method, and corrects the inaccurate re-scaling factor in the Kronecker-factored eigenbasis. Gao et al. (2020) considers a new approximation to the natural gradient, which approximates the Fisher information matrix (FIM) to a constant multiplied by the Kronecker product of two matrices and keeps the trace equal before and after the approximation. In this work, we combine the ideas of these two methods and propose Trace-restricted Eigenvalue-corrected Kronecker Factorization (TEKFAC). The proposed method not only corrects the inexact re-scaling factor under the Kronecker-factored eigenbasis, but also considers the new approximation method and the effective damping technique proposed in Gao et al. (2020). We also discuss the differences and relationships among the Kronecker-factored approximations. Empirically, our method outperforms SGD with momentum, Adam, EKFAC and TKFAC on several DNNs.

Kai-Xin Gao, Xiao-Lei Liu, Zheng-Hai Huang et al.

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. 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) in this work, 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.

Fan Yu, Zhuoyuan Yao, Xiong Wang et al.

Automatic speech recognition (ASR) has been significantly advanced with the use of deep learning and big data. However improving robustness, including achieving equally good performance on diverse speakers and accents, is still a challenging problem. In particular, the performance of children speech recognition (CSR) still lags behind due to 1) the speech and language characteristics of children's voice are substantially different from those of adults and 2) sizable open dataset for children speech is still not available in the research community. To address these problems, we launch the Children Speech Recognition Challenge (CSRC), as a flagship satellite event of IEEE SLT 2021 workshop. The challenge will release about 400 hours of Mandarin speech data for registered teams and set up two challenge tracks and provide a common testbed to benchmark the CSR performance. In this paper, we introduce the datasets, rules, evaluation method as well as baselines.

Zhenkun Cai, Kaihao Ma, Xiao Yan et al.

A good parallelization strategy can significantly improve the efficiency or reduce the cost for the distributed training of deep neural networks (DNNs). Recently, several methods have been proposed to find efficient parallelization strategies but they all optimize a single objective (e.g., execution time, memory consumption) and produce only one strategy. We propose FT, an efficient algorithm that searches for an optimal set of parallelization strategies to allow the trade-off among different objectives. FT can adapt to different scenarios by minimizing the memory consumption when the number of devices is limited and fully utilize additional resources to reduce the execution time. For popular DNN models (e.g., vision, language), an in-depth analysis is conducted to understand the trade-offs among different objectives and their influence on the parallelization strategies. We also develop a user-friendly system, called TensorOpt, which allows users to run their distributed DNN training jobs without caring the details of parallelization strategies. Experimental results show that FT runs efficiently and provides accurate estimation of runtime costs, and TensorOpt is more flexible in adapting to resource availability compared with existing frameworks.