Yuhong Guo

Jiale Xu, Rui Zhang, Cong Guo et al.

Large Language Models (LLMs) are widely used across various domains, processing millions of daily requests. This surge in demand poses significant challenges in optimizing throughput and latency while keeping costs manageable. The Key-Value (KV) cache, a standard method for retaining previous computations, makes LLM inference highly bounded by memory. While batching strategies can enhance performance, they frequently lead to significant memory fragmentation. Even though cutting-edge systems like vLLM mitigate KV cache fragmentation using paged Attention mechanisms, they still suffer from inefficient memory and computational operations due to the tightly coupled page management and computation kernels. This study introduces the vTensor, an innovative tensor structure for LLM inference based on GPU virtual memory management (VMM). vTensor addresses existing limitations by decoupling computation from memory defragmentation and offering dynamic extensibility. Our framework employs a CPU-GPU heterogeneous approach, ensuring efficient, fragmentation-free memory management while accommodating various computation kernels across different LLM architectures. Experimental results indicate that vTensor achieves an average speedup of 1.86x across different models, with up to 2.42x in multi-turn chat scenarios. Additionally, vTensor provides average speedups of 2.12x and 3.15x in kernel evaluation, reaching up to 3.92x and 3.27x compared to SGLang Triton prefix-prefilling kernels and vLLM paged Attention kernel, respectively. Furthermore, it frees approximately 71.25% (57GB) of memory on the NVIDIA A100 GPU compared to vLLM, enabling more memory-intensive workloads.

Qing En, Yuhong Guo

Medical image segmentation methods typically rely on numerous dense annotated images for model training, which are notoriously expensive and time-consuming to collect. To alleviate this burden, weakly supervised techniques have been exploited to train segmentation models with less expensive annotations. In this paper, we propose a novel point-supervised contrastive variance method (PSCV) for medical image semantic segmentation, which only requires one pixel-point from each organ category to be annotated. The proposed method trains the base segmentation network by using a novel contrastive variance (CV) loss to exploit the unlabeled pixels and a partial cross-entropy loss on the labeled pixels. The CV loss function is designed to exploit the statistical spatial distribution properties of organs in medical images and their variance distribution map representations to enforce discriminative predictions over the unlabeled pixels. Experimental results on two standard medical image datasets demonstrate that the proposed method outperforms the state-of-the-art weakly supervised methods on point-supervised medical image semantic segmentation tasks.

Qing En, Yuhong Guo

Medical image annotation typically requires expert knowledge and hence incurs time-consuming and expensive data annotation costs. To alleviate this burden, we propose a novel learning scenario, Exemplar Learning (EL), to explore automated learning processes for medical image segmentation with a single annotated image example. This innovative learning task is particularly suitable for medical image segmentation, where all categories of organs can be presented in one single image and annotated all at once. To address this challenging EL task, we propose an Exemplar Learning-based Synthesis Net (ELSNet) framework for medical image segmentation that enables innovative exemplar-based data synthesis, pixel-prototype based contrastive embedding learning, and pseudo-label based exploitation of the unlabeled data. Specifically, ELSNet introduces two new modules for image segmentation: an exemplar-guided synthesis module, which enriches and diversifies the training set by synthesizing annotated samples from the given exemplar, and a pixel-prototype based contrastive embedding module, which enhances the discriminative capacity of the base segmentation model via contrastive representation learning. Moreover, we deploy a two-stage process for segmentation model training, which exploits the unlabeled data with predicted pseudo segmentation labels. To evaluate this new learning framework, we conduct extensive experiments on several organ segmentation datasets and present an in-depth analysis. The empirical results show that the proposed exemplar learning framework produces effective segmentation results.

Hao Yan, Yuhong Guo

Unsupervised domain adaptation reduces the reliance on data annotation in deep learning by adapting knowledge from a source to a target domain. For privacy and efficiency concerns, source-free domain adaptation extends unsupervised domain adaptation by adapting a pre-trained source model to an unlabeled target domain without accessing the source data. However, most existing source-free domain adaptation methods to date focus on the transductive setting, where the target training set is also the testing set. In this paper, we address source-free domain adaptation in the more realistic inductive setting, where the target training and testing sets are mutually exclusive. We propose a new semi-supervised fine-tuning method named Dual Moving Average Pseudo-Labeling (DMAPL) for source-free inductive domain adaptation. We first split the unlabeled training set in the target domain into a pseudo-labeled confident subset and an unlabeled less-confident subset according to the prediction confidence scores from the pre-trained source model. Then we propose a soft-label moving-average updating strategy for the unlabeled subset based on a moving-average prototypical classifier, which gradually adapts the source model towards the target domain. Experiments show that our proposed method achieves state-of-the-art performance and outperforms previous methods by large margins.

Hanping Zhang, Yuhong Guo

As safety violations can lead to severe consequences in real-world robotic applications, the increasing deployment of Reinforcement Learning (RL) in robotic domains has propelled the study of safe exploration for reinforcement learning (safe RL). In this work, we propose a risk preventive training method for safe RL, which learns a statistical contrastive classifier to predict the probability of a state-action pair leading to unsafe states. Based on the predicted risk probabilities, we can collect risk preventive trajectories and reshape the reward function with risk penalties to induce safe RL policies. We conduct experiments in robotic simulation environments. The results show the proposed approach has comparable performance with the state-of-the-art model-based methods and outperforms conventional model-free safe RL approaches.

Marzi Heidari, Hanping Zhang, Yuhong Guo

In recent years, semi-supervised learning (SSL) has gained significant attention due to its ability to leverage both labeled and unlabeled data to improve model performance, especially when labeled data is scarce. However, most current SSL methods rely on heuristics or predefined rules for generating pseudo-labels and leveraging unlabeled data. They are limited to exploiting loss functions and regularization methods within the standard norm. In this paper, we propose a novel Reinforcement Learning (RL) Guided SSL method, RLGSSL, that formulates SSL as a one-armed bandit problem and deploys an innovative RL loss based on weighted reward to adaptively guide the learning process of the prediction model. RLGSSL incorporates a carefully designed reward function that balances the use of labeled and unlabeled data to enhance generalization performance. A semi-supervised teacher-student framework is further deployed to increase the learning stability. We demonstrate the effectiveness of RLGSSL through extensive experiments on several benchmark datasets and show that our approach achieves consistent superior performance compared to state-of-the-art SSL methods.

Hao Yan, Yuhong Guo

Domain generalization aims at developing suitable learning algorithms in source training domains such that the model learned can generalize well on a different unseen testing domain. We present a novel two-stage approach called Context-Aware Self-Adaptation (CASA) for domain generalization. CASA simulates an approximate meta-generalization scenario and incorporates a self-adaptation module to adjust pre-trained meta source models to the meta-target domains while maintaining their predictive capability on the meta-source domains. The core concept of self-adaptation involves leveraging contextual information, such as the mean of mini-batch features, as domain knowledge to automatically adapt a model trained in the first stage to new contexts in the second stage. Lastly, we utilize an ensemble of multiple meta-source models to perform inference on the testing domain. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on standard benchmarks.

Marzi Heidari, Yuhong Guo

Deep models trained on a single source domain often fail catastrophically under distribution shifts, a critical challenge in Single Domain Generalization (SDG). While existing methods focus on augmenting source data or learning invariant features, they neglect a readily available resource: textual descriptions of the target deployment environment. We propose Target-Oriented Single Domain Generalization (TO-SDG), a novel problem setup that leverages the textual description of the target domain, without requiring any target data, to guide model generalization. To address TO-SDG, we introduce Spectral TARget Alignment (STAR), a lightweight module that injects target semantics into source features by exploiting visual-language models (VLMs) such as CLIP. STAR uses a target-anchored subspace derived from the text embedding of the target description to recenter image features toward the deployment domain, then utilizes spectral projection to retain directions aligned with target cues while discarding source-specific noise. Moreover, we use a vision-language distillation to align backbone features with VLM's semantic geometry. STAR further employs feature-space Mixup to ensure smooth transitions between source and target-oriented representations. Experiments across various image classification and object detection benchmarks demonstrate STAR's superiority. This work establishes that minimal textual metadata, which is a practical and often overlooked resource, significantly enhances generalization under severe data constraints, opening new avenues for deploying robust models in target environments with unseen data.

Hanping Zhang, Yuhong Guo

Safe Reinforcement Learning (Safe RL) aims to ensure safety when an RL agent conducts learning by interacting with real-world environments where improper actions can induce high costs or lead to severe consequences. In this paper, we propose a novel Safe Skill Planning (SSkP) approach to enhance effective safe RL by exploiting auxiliary offline demonstration data. SSkP involves a two-stage process. First, we employ PU learning to learn a skill risk predictor from the offline demonstration data. Then, based on the learned skill risk predictor, we develop a novel risk planning process to enhance online safe RL and learn a risk-averse safe policy efficiently through interactions with the online RL environment, while simultaneously adapting the skill risk predictor to the environment. We conduct experiments in several benchmark robotic simulation environments. The experimental results demonstrate that the proposed approach consistently outperforms previous state-of-the-art safe RL methods.

Hanping Zhang, Yuhong Guo

Safe Reinforcement Learning (Safe RL) aims to train an RL agent to maximize its performance in real-world environments while adhering to safety constraints, as exceeding safety violation limits can result in severe consequences. In this paper, we propose a novel safe RL approach called Safety Modulated Policy Optimization (SMPO), which enables safe policy function learning within the standard policy optimization framework through safety modulated rewards. In particular, we consider safety violation costs as feedback from the RL environments that are parallel to the standard awards, and introduce a Q-cost function as safety critic to estimate expected future cumulative costs. Then we propose to modulate the rewards using a cost-aware weighting function, which is carefully designed to ensure the safety limits based on the estimation of the safety critic, while maximizing the expected rewards. The policy function and the safety critic are simultaneously learned through gradient descent during online interactions with the environment. We conduct experiments using multiple RL environments and the experimental results demonstrate that our method outperforms several classic and state-of-the-art comparison methods in terms of overall safe RL performance.

Hanping Zhang, Yuhong Guo

Learning from Demonstration (LfD) is a well-established problem in Reinforcement Learning (RL), which aims to facilitate rapid RL by leveraging expert demonstrations to pre-train the RL agent. However, the limited availability of expert demonstration data often hinders its ability to effectively aid downstream RL learning. To address this problem, we propose a novel two-stage method dubbed as Skill-enhanced Reinforcement Learning Acceleration (SeRLA). SeRLA introduces a skill-level adversarial Positive-Unlabeled (PU) learning model that extracts useful skill prior knowledge by learning from both expert demonstrations and general low-cost demonstrations in the offline prior learning stage. Building on this, it employs a skill-based soft actor-critic algorithm to leverage the acquired priors for efficient training of a skill policy network in the downstream online RL stage. In addition, we propose a simple skill-level data enhancement technique to mitigate data sparsity and further improve both skill prior learning and skill policy training. Experiments across multiple standard RL benchmarks demonstrate that SeRLA achieves state-of-the-art performance in accelerating reinforcement learning on downstream tasks, particularly in the early training phase.

Xuejun Han, Yuhong Guo

New objects are continuously emerging in the dynamically changing world and a real-world artificial intelligence system should be capable of continual and effectual adaptation to new emerging classes without forgetting old ones. In view of this, in this paper we tackle a challenging and practical continual learning scenario named few-shot class-incremental learning (FSCIL), in which labeled data are given for classes in a base session but very limited labeled instances are available for new incremental classes. To address this problem, we propose a novel and succinct approach by introducing deep dictionary learning which is a hybrid learning architecture that combines dictionary learning and visual representation learning to provide a better space for characterizing different classes. We simultaneously optimize the dictionary and the feature extraction backbone in the base session, while only finetune the dictionary in the incremental session for adaptation to novel classes, which can alleviate the forgetting on base classes compared to finetuning the entire model. To further facilitate future adaptation, we also incorporate multiple pseudo classes into the base session training so that certain space projected by dictionary can be reserved for future new concepts. The extensive experimental results on CIFAR100, miniImageNet and CUB200 validate the effectiveness of our approach compared to other SOTA methods.

Xuejun Han, Yuhong Guo

Classical machine learners are designed only to tackle one task without capability of adopting new emerging tasks or classes whereas such capacity is more practical and human-like in the real world. To address this shortcoming, continual machine learners are elaborated to commendably learn a stream of tasks with domain and class shifts among different tasks. In this paper, we propose a general feature-propagation based contrastive continual learning method which is capable of handling multiple continual learning scenarios. Specifically, we align the current and previous representation spaces by means of feature propagation and contrastive representation learning to bridge the domain shifts among distinct tasks. To further mitigate the class-wise shifts of the feature representation, a supervised contrastive loss is exploited to make the example embeddings of the same class closer than those of different classes. The extensive experimental results demonstrate the outstanding performance of the proposed method on six continual learning benchmarks compared to a group of cutting-edge continual learning methods.