Peng Sun

Pan Zhang, Xiaoyi Dong, Yuhang Zang et al. · pku

We present InternLM-XComposer-2.5 (IXC-2.5), a versatile large-vision language model that supports long-contextual input and output. IXC-2.5 excels in various text-image comprehension and composition applications, achieving GPT-4V level capabilities with merely 7B LLM backend. Trained with 24K interleaved image-text contexts, it can seamlessly extend to 96K long contexts via RoPE extrapolation. This long-context capability allows IXC-2.5 to excel in tasks requiring extensive input and output contexts. Compared to its previous 2.0 version, InternLM-XComposer-2.5 features three major upgrades in vision-language comprehension: (1) Ultra-High Resolution Understanding, (2) Fine-Grained Video Understanding, and (3) Multi-Turn Multi-Image Dialogue. In addition to comprehension, IXC-2.5 extends to two compelling applications using extra LoRA parameters for text-image composition: (1) Crafting Webpages and (2) Composing High-Quality Text-Image Articles. IXC-2.5 has been evaluated on 28 benchmarks, outperforming existing open-source state-of-the-art models on 16 benchmarks. It also surpasses or competes closely with GPT-4V and Gemini Pro on 16 key tasks. The InternLM-XComposer-2.5 is publicly available at https://github.com/InternLM/InternLM-XComposer.

Lei Sun, Daniel Gehrig, Christos Sakaridis et al.

Effective video frame interpolation hinges on the adept handling of motion in the input scene. Prior work acknowledges asynchronous event information for this, but often overlooks whether motion induces blur in the video, limiting its scope to sharp frame interpolation. We instead propose a unified framework for event-based frame interpolation that performs deblurring ad-hoc and thus works both on sharp and blurry input videos. Our model consists in a bidirectional recurrent network that incorporates the temporal dimension of interpolation and fuses information from the input frames and the events adaptively based on their temporal proximity. To enhance the generalization from synthetic data to real event cameras, we integrate self-supervised framework with the proposed model to enhance the generalization on real-world datasets in the wild. At the dataset level, we introduce a novel real-world high-resolution dataset with events and color videos named HighREV, which provides a challenging evaluation setting for the examined task. Extensive experiments show that our network consistently outperforms previous state-of-the-art methods on frame interpolation, single image deblurring, and the joint task of both. Experiments on domain transfer reveal that self-supervised training effectively mitigates the performance degradation observed when transitioning from synthetic data to real-world data. Code and datasets are available at https://github.com/AHupuJR/REFID.

Huali Zhou, Yueqian Lin, Yao Shi et al.

Although Singing Voice Synthesis (SVS) has made great strides with Text-to-Speech (TTS) techniques, multilingual singing voice modeling remains relatively unexplored. This paper presents BiSinger, a bilingual pop SVS system for English and Chinese Mandarin. Current systems require separate models per language and cannot accurately represent both Chinese and English, hindering code-switch SVS. To address this gap, we design a shared representation between Chinese and English singing voices, achieved by using the CMU dictionary with mapping rules. We fuse monolingual singing datasets with open-source singing voice conversion techniques to generate bilingual singing voices while also exploring the potential use of bilingual speech data. Experiments affirm that our language-independent representation and incorporation of related datasets enable a single model with enhanced performance in English and code-switch SVS while maintaining Chinese song performance. Audio samples are available at https://bisinger-svs.github.io.

Wei Gao, Qinghao Hu, Zhisheng Ye et al.

Deep learning (DL) shows its prosperity in a wide variety of fields. The development of a DL model is a time-consuming and resource-intensive procedure. Hence, dedicated GPU accelerators have been collectively constructed into a GPU datacenter. An efficient scheduler design for such GPU datacenter is crucially important to reduce the operational cost and improve resource utilization. However, traditional approaches designed for big data or high performance computing workloads can not support DL workloads to fully utilize the GPU resources. Recently, substantial schedulers are proposed to tailor for DL workloads in GPU datacenters. This paper surveys existing research efforts for both training and inference workloads. We primarily present how existing schedulers facilitate the respective workloads from the scheduling objectives and resource consumption features. Finally, we prospect several promising future research directions. More detailed summary with the surveyed paper and code links can be found at our project website: https://github.com/S-Lab-System-Group/Awesome-DL-Scheduling-Papers

Kun Yang, Jing Liu, Dingkang Yang et al.

With the rapid development of intelligent transportation system applications, a tremendous amount of multi-view video data has emerged to enhance vehicle perception. However, performing video analytics efficiently by exploiting the spatial-temporal redundancy from video data remains challenging. Accordingly, we propose a novel traffic-related framework named CEVAS to achieve efficient object detection using multi-view video data. Briefly, a fine-grained input filtering policy is introduced to produce a reasonable region of interest from the captured images. Also, we design a sharing object manager to manage the information of objects with spatial redundancy and share their results with other vehicles. We further derive a content-aware model selection policy to select detection methods adaptively. Experimental results show that our framework significantly reduces response latency while achieving the same detection accuracy as the state-of-the-art methods.

Zhiyuan Yan, Peng Sun, Yubo Lang et al.

Existing deepfake detectors face several challenges in achieving robustness and generalization. One of the primary reasons is their limited ability to extract relevant information from forgery videos, especially in the presence of various artifacts such as spatial, frequency, temporal, and landmark mismatches. Current detectors rely on pixel-level features that are easily affected by unknown disturbances or facial landmarks that do not provide sufficient information. Furthermore, most detectors cannot utilize information from multiple domains for detection, leading to limited effectiveness in identifying deepfake videos. To address these limitations, we propose a novel framework, namely Multimodal Graph Learning (MGL) that leverages information from multiple modalities using two GNNs and several multimodal fusion modules. At the frame level, we employ a bi-directional cross-modal transformer and an adaptive gating mechanism to combine the features from the spatial and frequency domains with the geometric-enhanced landmark features captured by a GNN. At the video level, we use a Graph Attention Network (GAT) to represent each frame in a video as a node in a graph and encode temporal information into the edges of the graph to extract temporal inconsistency between frames. Our proposed method aims to effectively identify and utilize distinguishing features for deepfake detection. We evaluate the effectiveness of our method through extensive experiments on widely-used benchmarks and demonstrate that our method outperforms the state-of-the-art detectors in terms of generalization ability and robustness against unknown disturbances.

Zepu Wang, Yuqi Nie, Peng Sun et al.

The criticality of prompt and precise traffic forecasting in optimizing traffic flow management in Intelligent Transportation Systems (ITS) has drawn substantial scholarly focus. Spatio-Temporal Graph Neural Networks (STGNNs) have been lauded for their adaptability to road graph structures. Yet, current research on STGNNs architectures often prioritizes complex designs, leading to elevated computational burdens with only minor enhancements in accuracy. To address this issue, we propose ST-MLP, a concise spatio-temporal model solely based on cascaded Multi-Layer Perceptron (MLP) modules and linear layers. Specifically, we incorporate temporal information, spatial information and predefined graph structure with a successful implementation of the channel-independence strategy - an effective technique in time series forecasting. Empirical results demonstrate that ST-MLP outperforms state-of-the-art STGNNs and other models in terms of accuracy and computational efficiency. Our finding encourages further exploration of more concise and effective neural network architectures in the field of traffic forecasting.

Bing Luo, Xiaomin Ouyang, Peng Sun et al.

With the rapid advancement of 5G networks, billions of smart Internet of Things (IoT) devices along with an enormous amount of data are generated at the network edge. While still at an early age, it is expected that the evolving 6G network will adopt advanced artificial intelligence (AI) technologies to collect, transmit, and learn this valuable data for innovative applications and intelligent services. However, traditional machine learning (ML) approaches require centralizing the training data in the data center or cloud, raising serious user-privacy concerns. Federated learning, as an emerging distributed AI paradigm with privacy-preserving nature, is anticipated to be a key enabler for achieving ubiquitous AI in 6G networks. However, there are several system and statistical heterogeneity challenges for effective and efficient FL implementation in 6G networks. In this article, we investigate the optimization approaches that can effectively address the challenging heterogeneity issues from three aspects: incentive mechanism design, network resource management, and personalized model optimization. We also present some open problems and promising directions for future research.

Guangliang Cheng, Peng Sun, Ting-Bing Xu et al.

Neural Architecture Search (NAS) has shown great potentials in automatically designing neural network architectures for real-time semantic segmentation. Unlike previous works that utilize a simplified search space with cell-sharing way, we introduce a new search space where a lightweight model can be more effectively searched by replacing the cell-sharing manner with cell-independent one. Based on this, the communication of local to global information is achieved through two well-designed modules. For local information exchange, a graph convolutional network (GCN) guided module is seamlessly integrated as a communication deliver between cells. For global information aggregation, we propose a novel dense-connected fusion module (cell) which aggregates long-range multi-level features in the network automatically. In addition, a latency-oriented constraint is endowed into the search process to balance the accuracy and latency. We name the proposed framework as Local-to-Global Information Communication Network Search (LGCNet). Extensive experiments on Cityscapes and CamVid datasets demonstrate that LGCNet achieves the new state-of-the-art trade-off between accuracy and speed. In particular, on Cityscapes dataset, LGCNet achieves the new best performance of 74.0\% mIoU with the speed of 115.2 FPS on Titan Xp.

Changnan Xiao, Yongxin Zhang, Xuefeng Huang et al.

Strategy card game is a well-known genre that is demanding on the intelligent game-play and can be an ideal test-bench for AI. Previous work combines an end-to-end policy function and an optimistic smooth fictitious play, which shows promising performances on the strategy card game Legend of Code and Magic. In this work, we apply such algorithms to Hearthstone, a famous commercial game that is more complicated in game rules and mechanisms. We further propose several improved techniques and consequently achieve significant progress. For a machine-vs-human test we invite a Hearthstone streamer whose best rank was top 10 of the official league in China region that is estimated to be of millions of players. Our models defeat the human player in all Best-of-5 tournaments of full games (including both deck building and battle), showing a strong capability of decision making.

Wei Xi, Yongxin Zhang, Changnan Xiao et al.

Deep Reinforcement Learning combined with Fictitious Play shows impressive results on many benchmark games, most of which are, however, single-stage. In contrast, real-world decision making problems may consist of multiple stages, where the observation spaces and the action spaces can be completely different across stages. We study a two-stage strategy card game Legends of Code and Magic and propose an end-to-end policy to address the difficulties that arise in multi-stage game. We also propose an optimistic smooth fictitious play algorithm to find the Nash Equilibrium for the two-player game. Our approach wins double championships of COG2022 competition. Extensive studies verify and show the advancement of our approach.

Trung Quoc Luong, Xinbo Zhang, Zhanming Jie et al.

One way to enhance the reasoning capability of Large Language Models (LLMs) is to conduct Supervised Fine-Tuning (SFT) using Chain-of-Thought (CoT) annotations. This approach does not show sufficiently strong generalization ability, however, because the training only relies on the given CoT data. In math problem-solving, for example, there is usually only one annotated reasoning path for each question in the training data. Intuitively, it would be better for the algorithm to learn from multiple annotated reasoning paths given a question. To address this issue, we propose a simple yet effective approach called Reinforced Fine-Tuning (ReFT) to enhance the generalizability of learning LLMs for reasoning, with math problem-solving as an example. ReFT first warmups the model with SFT, and then employs on-line reinforcement learning, specifically the PPO algorithm in this paper, to further fine-tune the model, where an abundance of reasoning paths are automatically sampled given the question and the rewards are naturally derived from the ground-truth answers. Extensive experiments on GSM8K, MathQA, and SVAMP datasets show that ReFT significantly outperforms SFT, and the performance can be potentially further boosted by combining inference-time strategies such as majority voting and re-ranking. Note that ReFT obtains the improvement by learning from the same training questions as SFT, without relying on extra or augmented training questions. This indicates a superior generalization ability for ReFT.

Kexuan Li, Susie Sinks, Peng Sun et al.

A bioequivalence study is a type of clinical trial designed to compare the biological equivalence of two different formulations of a drug. Such studies are typically conducted in controlled clinical settings with human subjects, who are randomly assigned to receive two formulations. The two formulations are then compared with respect to their pharmacokinetic profiles, which encompass the absorption, distribution, metabolism, and elimination of the drug. Under the guidance from Food and Drug Administration (FDA), for a size-$α$ bioequivalence test, the standard approach is to construct a $100(1-2α)\%$ confidence interval and verify if the confidence interval falls with the critical region. In this work, we clarify that $100(1-2α)\%$ confidence interval approach for bioequivalence testing yields a size-$α$ test only when the two one-sided tests in TOST are ``equal-tailed''. Furthermore, a $100(1-α)\%$ confidence interval approach is also discussed in the bioequivalence study.

Luyao Niu, Zepu Wang, Shuyi Guan et al.

While spatio-temporal Graph Neural Networks (GNNs) excel at modeling recurring traffic patterns, their reliability plummets during non-recurring events like accidents. This failure occurs because GNNs are fundamentally correlational models, learning historical patterns that are invalidated by the new causal factors introduced during disruptions. To address this, we propose Event-CausNet, a framework that uses a Large Language Model to quantify unstructured event reports, builds a causal knowledge base by estimating average treatment effects, and injects this knowledge into a dual-stream GNN-LSTM network using a novel causal attention mechanism to adjust and enhance the forecast. Experiments on a real-world dataset demonstrate that Event-CausNet achieves robust performance, reducing prediction error (MAE) by up to 35.87%, significantly outperforming state-of-the-art baselines. Our framework bridges the gap between correlational models and causal reasoning, providing a solution that is more accurate and transferable, while also offering crucial interpretability, providing a more reliable foundation for real-world traffic management during critical disruptions.

Shuohuan Wang, Yu Sun, Yang Xiang et al.

Pre-trained language models have achieved state-of-the-art results in various Natural Language Processing (NLP) tasks. GPT-3 has shown that scaling up pre-trained language models can further exploit their enormous potential. A unified framework named ERNIE 3.0 was recently proposed for pre-training large-scale knowledge enhanced models and trained a model with 10 billion parameters. ERNIE 3.0 outperformed the state-of-the-art models on various NLP tasks. In order to explore the performance of scaling up ERNIE 3.0, we train a hundred-billion-parameter model called ERNIE 3.0 Titan with up to 260 billion parameters on the PaddlePaddle platform. Furthermore, we design a self-supervised adversarial loss and a controllable language modeling loss to make ERNIE 3.0 Titan generate credible and controllable texts. To reduce the computation overhead and carbon emission, we propose an online distillation framework for ERNIE 3.0 Titan, where the teacher model will teach students and train itself simultaneously. ERNIE 3.0 Titan is the largest Chinese dense pre-trained model so far. Empirical results show that the ERNIE 3.0 Titan outperforms the state-of-the-art models on 68 NLP datasets.

Lei Sun, Christos Sakaridis, Jingyun Liang et al.

Traditional frame-based cameras inevitably suffer from motion blur due to long exposure times. As a kind of bio-inspired camera, the event camera records the intensity changes in an asynchronous way with high temporal resolution, providing valid image degradation information within the exposure time. In this paper, we rethink the eventbased image deblurring problem and unfold it into an end-to-end two-stage image restoration network. To effectively fuse event and image features, we design an event-image cross-modal attention module applied at multiple levels of our network, which allows to focus on relevant features from the event branch and filter out noise. We also introduce a novel symmetric cumulative event representation specifically for image deblurring as well as an event mask gated connection between the two stages of our network which helps avoid information loss. At the dataset level, to foster event-based motion deblurring and to facilitate evaluation on challenging real-world images, we introduce the Real Event Blur (REBlur) dataset, captured with an event camera in an illumination controlled optical laboratory. Our Event Fusion Network (EFNet) sets the new state of the art in motion deblurring, surpassing both the prior best-performing image-based method and all event-based methods with public implementations on the GoPro dataset (by up to 2.47dB) and on our REBlur dataset, even in extreme blurry conditions. The code and our REBlur dataset will be made publicly available.

Qinghao Hu, Peng Sun, Shengen Yan et al.

Modern GPU datacenters are critical for delivering Deep Learning (DL) models and services in both the research community and industry. When operating a datacenter, optimization of resource scheduling and management can bring significant financial benefits. Achieving this goal requires a deep understanding of the job features and user behaviors. We present a comprehensive study about the characteristics of DL jobs and resource management. First, we perform a large-scale analysis of real-world job traces from SenseTime. We uncover some interesting conclusions from the perspectives of clusters, jobs and users, which can facilitate the cluster system designs. Second, we introduce a general-purpose framework, which manages resources based on historical data. As case studies, we design: a Quasi-Shortest-Service-First scheduling service, which can minimize the cluster-wide average job completion time by up to 6.5x; and a Cluster Energy Saving service, which improves overall cluster utilization by up to 13%.

Peng Sun, Gabriel Draughon, Jerome Lynch

Coronavirus has been spreading around the world since the end of 2019. The virus can cause acute respiratory syndrome, which can be lethal, and is easily transmitted between hosts. Most states have issued state-at-home executive orders, however, parks and other public open spaces have largely remained open and are seeing sharp increases in public use. Therefore, in order to ensure public safety, it is imperative for patrons of public open spaces to practice safe hygiene and take preventative measures. This work provides a scalable sensing approach to detect physical activities within public open spaces and monitor adherence to social distancing guidelines suggested by the US Centers for Disease Control and Prevention (CDC). A deep learning-based computer vision sensing framework is designed to investigate the careful and proper utilization of parks and park facilities with hard surfaces (e.g. benches, fence poles, and trash cans) using video feeds from a pre-installed surveillance camera network. The sensing framework consists of a CNN-based object detector, a multi-target tracker, a mapping module, and a group reasoning module. The experiments are carried out during the COVID-19 pandemic between March 2020 and May 2020 across several key locations at the Detroit Riverfront Parks in Detroit, Michigan. The sensing framework is validated by comparing automatic sensing results with manually labeled ground-truth results. The proposed approach significantly improves the efficiency of providing spatial and temporal statistics of users in public open spaces by creating straightforward data visualizations for federal and state agencies. The results can also provide on-time triggering information for an alarming or actuator system which can later be added to intervene inappropriate behavior during this pandemic.

Peng Sun, Rui Hou, Jerome Lynch

Physical activities and social interactions are essential activities that ensure a healthy lifestyle. Public open spaces (POS), such as parks, plazas and greenways, are key environments that encourage those activities. To evaluate a POS, there is a need to study how humans use the facilities within it. However, traditional approaches to studying use of POS are manual and therefore time and labor intensive. They also may only provide qualitative insights. It is appealing to make use of surveillance cameras and to extract user-related information through computer vision. This paper proposes a proof-of-concept deep learning computer vision framework for measuring human activities quantitatively in POS and demonstrates a case study of the proposed framework using the Detroit Riverfront Conservancy (DRFC) surveillance camera network. A custom image dataset is presented to train the framework; the dataset includes 7826 fully annotated images collected from 18 cameras across the DRFC park space under various illumination conditions. Dataset analysis is also provided as well as a baseline model for one-step user localization and activity recognition. The mAP results are 77.5\% for {\it pedestrian} detection and 81.6\% for {\it cyclist} detection. Behavioral maps are autonomously generated by the framework to locate different POS users and the average error for behavioral localization is within 10 cm.

Tiezheng Ge, Liqin Zhao, Guorui Zhou et al.

In Taobao, the largest e-commerce platform in China, billions of items are provided and typically displayed with their images. For better user experience and business effectiveness, Click Through Rate (CTR) prediction in online advertising system exploits abundant user historical behaviors to identify whether a user is interested in a candidate ad. Enhancing behavior representations with user behavior images will help understand user's visual preference and improve the accuracy of CTR prediction greatly. So we propose to model user preference jointly with user behavior ID features and behavior images. However, training with user behavior images brings tens to hundreds of images in one sample, giving rise to a great challenge in both communication and computation. To handle these challenges, we propose a novel and efficient distributed machine learning paradigm called Advanced Model Server (AMS). With the well known Parameter Server (PS) framework, each server node handles a separate part of parameters and updates them independently. AMS goes beyond this and is designed to be capable of learning a unified image descriptor model shared by all server nodes which embeds large images into low dimensional high level features before transmitting images to worker nodes. AMS thus dramatically reduces the communication load and enables the arduous joint training process. Based on AMS, the methods of effectively combining the images and ID features are carefully studied, and then we propose a Deep Image CTR Model. Our approach is shown to achieve significant improvements in both online and offline evaluations, and has been deployed in Taobao display advertising system serving the main traffic.