Ji Liu

Guanchen Li, Yixing Xu, Zeping Li et al.

Structural pruning enhances hardware-agnostic inference efficiency for large language models (LLMs) yet often fails to maintain comparable performance. Local pruning performs efficient layer-by-layer compression but ignores global topology. Although global pruning aims to identify an optimal sparse model, intuitive methods typically adopt a two-stage paradigm that first evaluates substructure saliency and then applies global pruning, which ignores inter-structure dependencies and fails to achieve end-to-end optimization. To address these limitations, we propose Týr-the-Pruner, an efficient end-to-end search-based global structural pruning framework. This framework constructs a supernet by repeatedly applying local pruning across a range of sparsity ratios to each layer in an LLM, with the core goal of determining the optimal sparsity distribution under a target overall sparsity ratio. Concretely, we introduce an effective local pruning and an expectation error accumulation approach to improve supernet construction. Furthermore, we employ an iterative prune-and-search strategy with coarse-to-fine sparsity granularity to ensure efficient search convergence. Experimental results show that Týr-the-Pruner achieves state-of-the-art structural pruning, retaining 97% of the dense model's performance while removing a challenging 50% of Llama-3.1-70B's parameters. Code will be available at https://github.com/AMD-AGI/Tyr-the-Pruner.

Ji Liu, Zhihua Wu, Dianhai Yu et al.

Deep neural networks (DNNs) exploit many layers and a large number of parameters to achieve excellent performance. The training process of DNN models generally handles large-scale input data with many sparse features, which incurs high Input/Output (IO) cost, while some layers are compute-intensive. The training process generally exploits distributed computing resources to reduce training time. In addition, heterogeneous computing resources, e.g., CPUs, GPUs of multiple types, are available for the distributed training process. Thus, the scheduling of multiple layers to diverse computing resources is critical for the training process. To efficiently train a DNN model using the heterogeneous computing resources, we propose a distributed framework, i.e., Paddle-Heterogeneous Parameter Server (Paddle-HeterPS), composed of a distributed architecture and a Reinforcement Learning (RL)-based scheduling method. The advantages of Paddle-HeterPS are three-fold compared with existing frameworks. First, Paddle-HeterPS enables efficient training process of diverse workloads with heterogeneous computing resources. Second, Paddle-HeterPS exploits an RL-based method to efficiently schedule the workload of each layer to appropriate computing resources to minimize the cost while satisfying throughput constraints. Third, Paddle-HeterPS manages data storage and data communication among distributed computing resources. We carry out extensive experiments to show that Paddle-HeterPS significantly outperforms state-of-the-art approaches in terms of throughput (14.5 times higher) and monetary cost (312.3% smaller). The codes of the framework are publicly available at: https://github.com/PaddlePaddle/Paddle.

Haotao Wang, Tianlong Chen, Shupeng Gui et al.

Adversarial training and its many variants substantially improve deep network robustness, yet at the cost of compromising standard accuracy. Moreover, the training process is heavy and hence it becomes impractical to thoroughly explore the trade-off between accuracy and robustness. This paper asks this new question: how to quickly calibrate a trained model in-situ, to examine the achievable trade-offs between its standard and robust accuracies, without (re-)training it many times? Our proposed framework, Once-for-all Adversarial Training (OAT), is built on an innovative model-conditional training framework, with a controlling hyper-parameter as the input. The trained model could be adjusted among different standard and robust accuracies "for free" at testing time. As an important knob, we exploit dual batch normalization to separate standard and adversarial feature statistics, so that they can be learned in one model without degrading performance. We further extend OAT to a Once-for-all Adversarial Training and Slimming (OATS) framework, that allows for the joint trade-off among accuracy, robustness and runtime efficiency. Experiments show that, without any re-training nor ensembling, OAT/OATS achieve similar or even superior performance compared to dedicatedly trained models at various configurations. Our codes and pretrained models are available at: https://github.com/VITA-Group/Once-for-All-Adversarial-Training.

Xichuan Zhou, Kui Liu, Cong Shi et al.

Recent researches on information bottleneck shed new light on the continuous attempts to open the black box of neural signal encoding. Inspired by the problem of lossy signal compression for wireless communication, this paper presents a Bitwise Information Bottleneck approach for quantizing and encoding neural network activations. Based on the rate-distortion theory, the Bitwise Information Bottleneck attempts to determine the most significant bits in activation representation by assigning and approximating the sparse coefficient associated with each bit. Given the constraint of a limited average code rate, the information bottleneck minimizes the rate-distortion for optimal activation quantization in a flexible layer-by-layer manner. Experiments over ImageNet and other datasets show that, by minimizing the quantization rate-distortion of each layer, the neural network with information bottlenecks achieves the state-of-the-art accuracy with low-precision activation. Meanwhile, by reducing the code rate, the proposed method can improve the memory and computational efficiency by over six times compared with the deep neural network with standard single-precision representation. Codes will be available on GitHub when the paper is accepted \url{https://github.com/BitBottleneck/PublicCode}.

Ji Liu, Lei Zhang

Recently, learning to hash has been widely studied for image retrieval thanks to the computation and storage efficiency of binary codes. For most existing learning to hash methods, sufficient training images are required and used to learn precise hashing codes. However, in some real-world applications, there are not always sufficient training images in the domain of interest. In addition, some existing supervised approaches need a amount of labeled data, which is an expensive process in term of time, label and human expertise. To handle such problems, inspired by transfer learning, we propose a simple yet effective unsupervised hashing method named Optimal Projection Guided Transfer Hashing (GTH) where we borrow the images of other different but related domain i.e., source domain to help learn precise hashing codes for the domain of interest i.e., target domain. Besides, we propose to seek for the maximum likelihood estimation (MLE) solution of the hashing functions of target and source domains due to the domain gap. Furthermore,an alternating optimization method is adopted to obtain the two projections of target and source domains such that the domain hashing disparity is reduced gradually. Extensive experiments on various benchmark databases verify that our method outperforms many state-of-the-art learning to hash methods. The implementation details are available at https://github.com/liuji93/GTH.

Juncheng Jia, Ji Liu, Chao Huo et al.

Federated Learning (FL) has emerged as a compelling methodology for the management of distributed data, marked by significant advancements in recent years. In this paper, we propose an efficient FL approach that capitalizes on additional downlink bandwidth resources to ensure timely update dissemination. Initially, we implement this strategy within an asynchronous framework, introducing the Asynchronous Staleness-aware Model Update (FedASMU), which integrates both server-side and device-side methodologies. On the server side, we present an asynchronous FL system model that employs a dynamic model aggregation technique, which harmonizes local model updates with the global model to enhance both accuracy and efficiency. Concurrently, on the device side, we propose an adaptive model adjustment mechanism that integrates the latest global model with local models during training to further elevate accuracy. Subsequently, we extend this approach to a synchronous context, referred to as FedSSMU. Theoretical analyses substantiate the convergence of our proposed methodologies. Extensive experiments, encompassing six models and five public datasets, demonstrate that FedASMU and FedSSMU significantly surpass baseline methods in terms of both accuracy (up to 145.87%) and efficiency (up to 97.59%).

Zeping Li, Xinlong Yang, Ziheng Gao et al.

Large Language Models (LLMs) inherently use autoregressive decoding, which lacks parallelism in inference and results in significantly slow inference speed. While methods such as Medusa constructs parallelized heads, they lack adequate information interaction across different prediction positions. To overcome this limitation, we introduce Amphista, an enhanced speculative decoding framework that builds upon Medusa. Specifically, Amphista models an Auto-embedding Block capable of parallel inference, incorporating bi-directional attention to enable interaction between different drafting heads. Additionally, Amphista integrates Staged Adaptation Layers, which ensure a seamless transition of semantic information from the target model's autoregressive inference to the drafting heads' non-autoregressive inference, effectively achieving paradigm shift and feature fusion. Experimental results on Vicuna models using MT-Bench and Spec-Bench demonstrate that Amphista achieves substantial acceleration while maintaining generation quality. On MT-Bench, Amphista delivers up to 2.75$\times$ speedup over vanilla autoregressive decoding and 1.40$\times$ over Medusa on Vicuna 33B in wall-clock time.

Ji Liu, Ce Zhang

Scalable and efficient distributed learning is one of the main driving forces behind the recent rapid advancement of machine learning and artificial intelligence. One prominent feature of this topic is that recent progresses have been made by researchers in two communities: (1) the system community such as database, data management, and distributed systems, and (2) the machine learning and mathematical optimization community. The interaction and knowledge sharing between these two communities has led to the rapid development of new distributed learning systems and theory. In this work, we hope to provide a brief introduction of some distributed learning techniques that have recently been developed, namely lossy communication compression (e.g., quantization and sparsification), asynchronous communication, and decentralized communication. One special focus in this work is on making sure that it can be easily understood by researchers in both communities -- On the system side, we rely on a simplified system model hiding many system details that are not necessary for the intuition behind the system speedups; while, on the theory side, we rely on minimal assumptions and significantly simplify the proof of some recent work to achieve comparable results.

Ji Liu, Hang Zhao, Jiyuan Yang et al.

Given a large number of online services on the Internet, from time to time, people are still struggling to find out the services that they need. On the other hand, when there are considerable research and development on service discovery and service recommendation, most of the related work are centralized and thus suffers inherent shortages of the centralized systems, e.g., adv-driven, lack at trust, transparence and fairness. In this paper, we propose a ServiceNet - a peer-to-peer (P2P) service network for service discovery and service recommendation. ServiceNet is inspired by blockchain technology and aims at providing an open, transparent and self-growth, and self-management service ecosystem. The paper will present the basic idea, an architecture design of the prototype, and an initial implementation and performance evaluation the prototype design.

Yi Guo, Ji Liu

Guided sparse depth upsampling aims to upsample an irregularly sampled sparse depth map when an aligned high-resolution color image is given as guidance. Many neural networks have been designed for this task. However, they often ignore the structural difference between the depth and the color image, resulting in obvious artifacts such as texture copy and depth blur at the upsampling depth. Inspired by the normalized convolution operation, we propose a guided convolutional layer to recover dense depth from sparse and irregular depth image with an depth edge image as guidance. Our novel guided network can prevent the depth value from crossing the depth edge to facilitate upsampling. We further design a convolution network based on proposed convolutional layer to combine the advantages of different algorithms and achieve better performance. We conduct comprehensive experiments to verify our method on real-world indoor and synthetic outdoor datasets. Our method produces strong results. It outperforms state-of-the-art methods on the Virtual KITTI dataset and the Middlebury dataset. It also presents strong generalization capability under different 3D point densities, various lighting and weather conditions.

Hui Chen, Guiguang Ding, Xudong Liu et al.

Enabling bi-directional retrieval of images and texts is important for understanding the correspondence between vision and language. Existing methods leverage the attention mechanism to explore such correspondence in a fine-grained manner. However, most of them consider all semantics equally and thus align them uniformly, regardless of their diverse complexities. In fact, semantics are diverse (i.e. involving different kinds of semantic concepts), and humans usually follow a latent structure to combine them into understandable languages. It may be difficult to optimally capture such sophisticated correspondences in existing methods. In this paper, to address such a deficiency, we propose an Iterative Matching with Recurrent Attention Memory (IMRAM) method, in which correspondences between images and texts are captured with multiple steps of alignments. Specifically, we introduce an iterative matching scheme to explore such fine-grained correspondence progressively. A memory distillation unit is used to refine alignment knowledge from early steps to later ones. Experiment results on three benchmark datasets, i.e. Flickr8K, Flickr30K, and MS COCO, show that our IMRAM achieves state-of-the-art performance, well demonstrating its effectiveness. Experiments on a practical business advertisement dataset, named \Ads{}, further validates the applicability of our method in practical scenarios.

Yan Lin, Ji Liu, Jianlin Zhou

Curve skeleton extraction from unorganized point cloud is a fundamental task of computer vision and three-dimensional data preprocessing and visualization. A great amount of work has been done to extract skeleton from point cloud. but the lack of standard datasets of point cloud with ground truth skeleton makes it difficult to evaluate these algorithms. In this paper, we construct a brand new tree-structured point cloud dataset, including ground truth skeletons, and point cloud models. In addition, four types of point cloud are built on clean point cloud: point clouds with noise, point clouds with missing data, point clouds with different density, and point clouds with uneven density distribution. We first use tree editor to build the tree skeleton and corresponding mesh model. Since the implicit surface is sufficiently expressive to retain the edges and details of the complex branches model, we use the implicit surface to model the triangular mesh. With the implicit surface, virtual scanner is applied to the sampling of point cloud. Finally, considering the challenges in skeleton extraction, we introduce different methods to build four different types of point cloud models. This dataset can be used as standard dataset for skeleton extraction algorithms. And the evaluation between skeleton extraction algorithms can be performed by comparing the ground truth skeleton with the extracted skeleton.

Xiangru Lian, Ji Liu

Batch Normalization (BN) has been used extensively in deep learning to achieve faster training process and better resulting models. However, whether BN works strongly depends on how the batches are constructed during training and it may not converge to a desired solution if the statistics on a batch are not close to the statistics over the whole dataset. In this paper, we try to understand BN from an optimization perspective by formulating the optimization problem which motivates BN. We show when BN works and when BN does not work by analyzing the optimization problem. We then propose a refinement of BN based on compositional optimization techniques called Full Normalization (FN) to alleviate the issues of BN when the batches are not constructed ideally. We provide convergence analysis for FN and empirically study its effectiveness to refine BN.

Yue Wu, Youzuo Lin, Zheng Zhou et al.

Automatic event detection from time series signals has wide applications, such as abnormal event detection in video surveillance and event detection in geophysical data. Traditional detection methods detect events primarily by the use of similarity and correlation in data. Those methods can be inefficient and yield low accuracy. In recent years, because of the significantly increased computational power, machine learning techniques have revolutionized many science and engineering domains. In this study, we apply a deep-learning-based method to the detection of events from time series seismic signals. However, a direct adaptation of the similar ideas from 2D object detection to our problem faces two challenges. The first challenge is that the duration of earthquake event varies significantly; The other is that the proposals generated are temporally correlated. To address these challenges, we propose a novel cascaded region-based convolutional neural network to capture earthquake events in different sizes, while incorporating contextual information to enrich features for each individual proposal. To achieve a better generalization performance, we use densely connected blocks as the backbone of our network. Because of the fact that some positive events are not correctly annotated, we further formulate the detection problem as a learning-from-noise problem. To verify the performance of our detection methods, we employ our methods to seismic data generated from a bi-axial "earthquake machine" located at Rock Mechanics Laboratory, and we acquire labels with the help of experts. Through our numerical tests, we show that our novel detection techniques yield high accuracy. Therefore, our novel deep-learning-based detection methods can potentially be powerful tools for locating events from time series data in various applications.

Gan Sun, Yang Cong, Ji Liu et al.

The state-of-the-art online learning approaches are only capable of learning the metric for predefined tasks. In this paper, we consider lifelong learning problem to mimic "human learning", i.e., endowing a new capability to the learned metric for a new task from new online samples and incorporating previous experiences and knowledge. Therefore, we propose a new metric learning framework: lifelong metric learning (LML), which only utilizes the data of the new task to train the metric model while preserving the original capabilities. More specifically, the proposed LML maintains a common subspace for all learned metrics, named lifelong dictionary, transfers knowledge from the common subspace to each new metric task with task-specific idiosyncrasy, and redefines the common subspace over time to maximize performance across all metric tasks. For model optimization, we apply online passive aggressive optimization algorithm to solve the proposed LML framework, where the lifelong dictionary and task-specific partition are optimized alternatively and consecutively. Finally, we evaluate our approach by analyzing several multi-task metric learning datasets. Extensive experimental results demonstrate effectiveness and efficiency of the proposed framework.

Mengdi Wang, Ji Liu, Ethan X. Fang

Consider the stochastic composition optimization problem where the objective is a composition of two expected-value functions. We propose a new stochastic first-order method, namely the accelerated stochastic compositional proximal gradient (ASC-PG) method, which updates based on queries to the sampling oracle using two different timescales. The ASC-PG is the first proximal gradient method for the stochastic composition problem that can deal with nonsmooth regularization penalty. We show that the ASC-PG exhibits faster convergence than the best known algorithms, and that it achieves the optimal sample-error complexity in several important special cases. We further demonstrate the application of ASC-PG to reinforcement learning and conduct numerical experiments.