Unlike existing knowledge distillation methods focus on the baseline settings, where the teacher models and training strategies are not that strong and competing as state-of-the-art approaches, this paper presents a method dubbed DIST to distill better from a stronger teacher. We empirically find that the discrepancy of predictions between the student and a stronger teacher may tend to be fairly severer. As a result, the exact match of predictions in KL divergence would disturb the training and make existing methods perform poorly. In this paper, we show that simply preserving the relations between the predictions of teacher and student would suffice, and propose a correlation-based loss to capture the intrinsic inter-class relations from the teacher explicitly. Besides, considering that different instances have different semantic similarities to each class, we also extend this relational match to the intra-class level. Our method is simple yet practical, and extensive experiments demonstrate that it adapts well to various architectures, model sizes and training strategies, and can achieve state-of-the-art performance consistently on image classification, object detection, and semantic segmentation tasks. Code is available at: https://github.com/hunto/DIST_KD .
Vision transformers (ViTs) are usually considered to be less light-weight than convolutional neural networks (CNNs) due to the lack of inductive bias. Recent works thus resort to convolutions as a plug-and-play module and embed them in various ViT counterparts. In this paper, we argue that the convolutional kernels perform information aggregation to connect all tokens; however, they would be actually unnecessary for light-weight ViTs if this explicit aggregation could function in a more homogeneous way. Inspired by this, we present LightViT as a new family of light-weight ViTs to achieve better accuracy-efficiency balance upon the pure transformer blocks without convolution. Concretely, we introduce a global yet efficient aggregation scheme into both self-attention and feed-forward network (FFN) of ViTs, where additional learnable tokens are introduced to capture global dependencies; and bi-dimensional channel and spatial attentions are imposed over token embeddings. Experiments show that our model achieves significant improvements on image classification, object detection, and semantic segmentation tasks. For example, our LightViT-T achieves 78.7% accuracy on ImageNet with only 0.7G FLOPs, outperforming PVTv2-B0 by 8.2% while 11% faster on GPU. Code is available at https://github.com/hunto/LightViT.
Distilling from the feature maps can be fairly effective for dense prediction tasks since both the feature discriminability and localization priors can be well transferred. However, not every pixel contributes equally to the performance, and a good student should learn from what really matters to the teacher. In this paper, we introduce a learnable embedding dubbed receptive token to localize those pixels of interests (PoIs) in the feature map, with a distillation mask generated via pixel-wise attention. Then the distillation will be performed on the mask via pixel-wise reconstruction. In this way, a distillation mask actually indicates a pattern of pixel dependencies within feature maps of teacher. We thus adopt multiple receptive tokens to investigate more sophisticated and informative pixel dependencies to further enhance the distillation. To obtain a group of masks, the receptive tokens are learned via the regular task loss but with teacher fixed, and we also leverage a Dice loss to enrich the diversity of learned masks. Our method dubbed MasKD is simple and practical, and needs no priors of tasks in application. Experiments show that our MasKD can achieve state-of-the-art performance consistently on object detection and semantic segmentation benchmarks. Code is available at: https://github.com/hunto/MasKD .
To break the bottlenecks of mainstream cloud-based machine learning (ML) paradigm, we adopt device-cloud collaborative ML and build the first end-to-end and general-purpose system, called Walle, as the foundation. Walle consists of a deployment platform, distributing ML tasks to billion-scale devices in time; a data pipeline, efficiently preparing task input; and a compute container, providing a cross-platform and high-performance execution environment, while facilitating daily task iteration. Specifically, the compute container is based on Mobile Neural Network (MNN), a tensor compute engine along with the data processing and model execution libraries, which are exposed through a refined Python thread-level virtual machine (VM) to support diverse ML tasks and concurrent task execution. The core of MNN is the novel mechanisms of operator decomposition and semi-auto search, sharply reducing the workload in manually optimizing hundreds of operators for tens of hardware backends and further quickly identifying the best backend with runtime optimization for a computation graph. The data pipeline introduces an on-device stream processing framework to enable processing user behavior data at source. The deployment platform releases ML tasks with an efficient push-then-pull method and supports multi-granularity deployment policies. We evaluate Walle in practical e-commerce application scenarios to demonstrate its effectiveness, efficiency, and scalability. Extensive micro-benchmarks also highlight the superior performance of MNN and the Python thread-level VM. Walle has been in large-scale production use in Alibaba, while MNN has been open source with a broad impact in the community.
Task automation of surgical robot has the potentials to improve surgical efficiency. Recent reinforcement learning (RL) based approaches provide scalable solutions to surgical automation, but typically require extensive data collection to solve a task if no prior knowledge is given. This issue is known as the exploration challenge, which can be alleviated by providing expert demonstrations to an RL agent. Yet, how to make effective use of demonstration data to improve exploration efficiency still remains an open challenge. In this work, we introduce Demonstration-guided EXploration (DEX), an efficient reinforcement learning algorithm that aims to overcome the exploration problem with expert demonstrations for surgical automation. To effectively exploit demonstrations, our method estimates expert-like behaviors with higher values to facilitate productive interactions, and adopts non-parametric regression to enable such guidance at states unobserved in demonstration data. Extensive experiments on $10$ surgical manipulation tasks from SurRoL, a comprehensive surgical simulation platform, demonstrate significant improvements in the exploration efficiency and task success rates of our method. Moreover, we also deploy the learned policies to the da Vinci Research Kit (dVRK) platform to show the effectiveness on the real robot. Code is available at https://github.com/med-air/DEX.
Multi-label classification (MLC) faces challenges from label noise in training data due to annotating diverse semantic labels for each image. Current methods mainly target identifying and correcting label mistakes using trained MLC models, but still struggle with persistent noisy labels during training, resulting in imprecise recognition and reduced performance. Our paper addresses label noise in MLC by introducing a positive and unlabeled multi-label classification (PU-MLC) method. To counteract noisy labels, we directly discard negative labels, focusing on the abundance of negative labels and the origin of most noisy labels. PU-MLC employs positive-unlabeled learning, training the model with only positive labels and unlabeled data. The method incorporates adaptive re-balance factors and temperature coefficients in the loss function to address label distribution imbalance and prevent over-smoothing of probabilities during training. Additionally, we introduce a local-global convolution module to capture both local and global dependencies in the image without requiring backbone retraining. PU-MLC proves effective on MLC and MLC with partial labels (MLC-PL) tasks, demonstrating significant improvements on MS-COCO and PASCAL VOC datasets with fewer annotations. Code is available at: https://github.com/TAKELAMAG/PU-MLC.
Motion artifacts compromise the quality of magnetic resonance imaging (MRI) and pose challenges to achieving diagnostic outcomes and image-guided therapies. In recent years, supervised deep learning approaches have emerged as successful solutions for motion artifact reduction (MAR). One disadvantage of these methods is their dependency on acquiring paired sets of motion artifact-corrupted (MA-corrupted) and motion artifact-free (MA-free) MR images for training purposes. Obtaining such image pairs is difficult and therefore limits the application of supervised training. In this paper, we propose a novel UNsupervised Abnormality Extraction Network (UNAEN) to alleviate this problem. Our network is capable of working with unpaired MA-corrupted and MA-free images. It converts the MA-corrupted images to MA-reduced images by extracting abnormalities from the MA-corrupted images using a proposed artifact extractor, which intercepts the residual artifact maps from the MA-corrupted MR images explicitly, and a reconstructor to restore the original input from the MA-reduced images. The performance of UNAEN was assessed by experimenting with various publicly available MRI datasets and comparing them with state-of-the-art methods. The quantitative evaluation demonstrates the superiority of UNAEN over alternative MAR methods and visually exhibits fewer residual artifacts. Our results substantiate the potential of UNAEN as a promising solution applicable in real-world clinical environments, with the capability to enhance diagnostic accuracy and facilitate image-guided therapies. Our codes are publicly available at https://github.com/YuSheng-Zhou/UNAEN.
As an emerging technology and a relatively affordable device, the 4D imaging radar has already been confirmed effective in performing 3D object detection in autonomous driving. Nevertheless, the sparsity and noisiness of 4D radar point clouds hinder further performance improvement, and in-depth studies about its fusion with other modalities are lacking. On the other hand, as a new image view transformation strategy, "sampling" has been applied in a few image-based detectors and shown to outperform the widely applied "depth-based splatting" proposed in Lift-Splat-Shoot (LSS), even without image depth prediction. However, the potential of "sampling" is not fully unleashed. This paper investigates the "sampling" view transformation strategy on the camera and 4D imaging radar fusion-based 3D object detection. LiDAR Excluded Lean (LXL) model, predicted image depth distribution maps and radar 3D occupancy grids are generated from image perspective view (PV) features and radar bird's eye view (BEV) features, respectively. They are sent to the core of LXL, called "radar occupancy-assisted depth-based sampling", to aid image view transformation. We demonstrated that more accurate view transformation can be performed by introducing image depths and radar information to enhance the "sampling" strategy. Experiments on VoD and TJ4DRadSet datasets show that the proposed method outperforms the state-of-the-art 3D object detection methods by a significant margin without bells and whistles. Ablation studies demonstrate that our method performs the best among different enhancement settings.
Multi-object tracking (MOT) is among crucial applications in modern advanced driver assistance systems (ADAS) and autonomous driving (AD) systems. The global nearest neighbor (GNN) filter, as the earliest random vector-based Bayesian tracking framework, has been adopted in most of state-of-the-arts trackers in the automotive industry. The development of random finite set (RFS) theory facilitates a mathematically rigorous treatment of the MOT problem, and different variants of RFS-based Bayesian filters have then been proposed. However, their effectiveness in the real ADAS and AD application is still an open problem. In this paper, it is demonstrated that the latest RFS-based Bayesian tracking framework could be superior to typical random vector-based Bayesian tracking framework via a systematic comparative study of both traditional random vector-based Bayesian filters with rule-based heuristic track maintenance and RFS-based Bayesian filters on the nuScenes validation dataset. An RFS-based tracker, namely Poisson multi-Bernoulli filter using the global nearest neighbor (GNN-PMB), is proposed to LiDAR-based MOT tasks. This GNN-PMB tracker is simple to use, and it achieves competitive results on the nuScenes dataset. Specifically, the proposed GNN-PMB tracker outperforms most state-of-the-art LiDAR-only trackers and LiDAR and camera fusion-based trackers, ranking the $3^{rd}$ among all LiDAR-only trackers on nuScenes 3D tracking challenge leader board at the time of submission.
The 4D Millimeter wave (mmWave) radar is a promising technology for vehicle sensing due to its cost-effectiveness and operability in adverse weather conditions. However, the adoption of this technology has been hindered by sparsity and noise issues in radar point cloud data. This paper introduces spatial multi-representation fusion (SMURF), a novel approach to 3D object detection using a single 4D imaging radar. SMURF leverages multiple representations of radar detection points, including pillarization and density features of a multi-dimensional Gaussian mixture distribution through kernel density estimation (KDE). KDE effectively mitigates measurement inaccuracy caused by limited angular resolution and multi-path propagation of radar signals. Additionally, KDE helps alleviate point cloud sparsity by capturing density features. Experimental evaluations on View-of-Delft (VoD) and TJ4DRadSet datasets demonstrate the effectiveness and generalization ability of SMURF, outperforming recently proposed 4D imaging radar-based single-representation models. Moreover, while using 4D imaging radar only, SMURF still achieves comparable performance to the state-of-the-art 4D imaging radar and camera fusion-based method, with an increase of 1.22% in the mean average precision on bird's-eye view of TJ4DRadSet dataset and 1.32% in the 3D mean average precision on the entire annotated area of VoD dataset. Our proposed method demonstrates impressive inference time and addresses the challenges of real-time detection, with the inference time no more than 0.05 seconds for most scans on both datasets. This research highlights the benefits of 4D mmWave radar and is a strong benchmark for subsequent works regarding 3D object detection with 4D imaging radar.
In autonomous driving, LiDAR and radar are crucial for environmental perception. LiDAR offers precise 3D spatial sensing information but struggles in adverse weather like fog. Conversely, radar signals can penetrate rain or mist due to their specific wavelength but are prone to noise disturbances. Recent state-of-the-art works reveal that the fusion of radar and LiDAR can lead to robust detection in adverse weather. The existing works adopt convolutional neural network architecture to extract features from each sensor data, then align and aggregate the two branch features to predict object detection results. However, these methods have low accuracy of predicted bounding boxes due to a simple design of label assignment and fusion strategies. In this paper, we propose a bird's-eye view fusion learning-based anchor box-free object detection system, which fuses the feature derived from the radar range-azimuth heatmap and the LiDAR point cloud to estimate possible objects. Different label assignment strategies have been designed to facilitate the consistency between the classification of foreground or background anchor points and the corresponding bounding box regressions. Furthermore, the performance of the proposed object detector is further enhanced by employing a novel interactive transformer module. The superior performance of the methods proposed in this paper has been demonstrated using the recently published Oxford Radar RobotCar dataset. Our system's average precision significantly outperforms the state-of-the-art method by 13.1% and 19.0% at Intersection of Union (IoU) of 0.8 under 'Clear+Foggy' training conditions for 'Clear' and 'Foggy' testing, respectively.
We introduce TACO, an open-source, large-scale code generation dataset, with a focus on the optics of algorithms, designed to provide a more challenging training dataset and evaluation benchmark in the field of code generation models. TACO includes competition-level programming questions that are more challenging, to enhance or evaluate problem understanding and reasoning abilities in real-world programming scenarios. There are 25433 and 1000 coding problems in training and test set, as well as up to 1.55 million diverse solution answers. Moreover, each TACO problem includes several fine-grained labels such as task topics, algorithms, programming skills, and difficulty levels, providing a more precise reference for the training and evaluation of code generation models. The dataset and evaluation scripts are available on Hugging Face Hub (https://huggingface.co/datasets/BAAI/TACO) and Github (https://github.com/FlagOpen/TACO).
High-resolution (HR) magnetic resonance imaging is critical in aiding doctors in their diagnoses and image-guided treatments. However, acquiring HR images can be time-consuming and costly. Consequently, deep learning-based super-resolution reconstruction (SRR) has emerged as a promising solution for generating super-resolution (SR) images from low-resolution (LR) images. Unfortunately, training such neural networks requires aligned authentic HR and LR image pairs, which are challenging to obtain due to patient movements during and between image acquisitions. While rigid movements of hard tissues can be corrected with image registration, aligning deformed soft tissues is complex, making it impractical to train neural networks with authentic HR and LR image pairs. Previous studies have focused on SRR using authentic HR images and down-sampled synthetic LR images. However, the difference in degradation representations between synthetic and authentic LR images suppresses the quality of SR images reconstructed from authentic LR images. To address this issue, we propose a novel Unsupervised Degradation Adaptation Network (UDEAN). Our network consists of a degradation learning network and an SRR network. The degradation learning network downsamples the HR images using the degradation representation learned from the misaligned or unpaired LR images. The SRR network then learns the mapping from the down-sampled HR images to the original ones. Experimental results show that our method outperforms state-of-the-art networks and is a promising solution to the challenges in clinical settings.
Issue resolution has made remarkable progress thanks to the advanced reasoning capabilities of large language models (LLMs). Recently, agent-based frameworks such as SWE-agent have further advanced this progress by enabling autonomous, tool-using agents to tackle complex software engineering tasks. While existing agent-based issue resolution approaches are primarily based on agents' independent explorations, they often get stuck in local solutions and fail to identify issue patterns that span across different parts of the codebase. To address this limitation, we propose SWE-Debate, a competitive multi-agent debate framework that encourages diverse reasoning paths and achieves more consolidated issue localization. SWE-Debate first creates multiple fault propagation traces as localization proposals by traversing a code dependency graph. Then, it organizes a three-round debate among specialized agents, each embodying distinct reasoning perspectives along the fault propagation trace. This structured competition enables agents to collaboratively converge on a consolidated fix plan. Finally, this consolidated fix plan is integrated into an MCTS-based code modification agent for patch generation. Experiments on the SWE-bench benchmark show that SWE-Debate achieves new state-of-the-art results in open-source agent frameworks and outperforms baselines by a large margin.
Knowledge distillation (KD) has proven highly effective for compressing large models and enhancing the performance of smaller ones. However, its effectiveness diminishes in cross-modal scenarios, such as vision-to-language distillation, where inconsistencies in representation across modalities lead to difficult knowledge transfer. To address this challenge, we propose frequency-decoupled cross-modal knowledge distillation, a method designed to decouple and balance knowledge transfer across modalities by leveraging frequency-domain features. We observed that low-frequency features exhibit high consistency across different modalities, whereas high-frequency features demonstrate extremely low cross-modal similarity. Accordingly, we apply distinct losses to these features: enforcing strong alignment in the low-frequency domain and introducing relaxed alignment for high-frequency features. We also propose a scale consistency loss to address distributional shifts between modalities, and employ a shared classifier to unify feature spaces. Extensive experiments across multiple benchmark datasets show our method substantially outperforms traditional KD and state-of-the-art cross-modal KD approaches. Code is available at https://github.com/Johumliu/FD-CMKD.
Despite the remarkable multimodal capabilities of Large Vision-Language Models (LVLMs), discrepancies often occur between visual inputs and textual outputs--a phenomenon we term visual hallucination. This critical reliability gap poses substantial risks in safety-critical Artificial Intelligence (AI) applications, necessitating a comprehensive evaluation benchmark and effective detection methods. Firstly, we observe that existing visual-centric hallucination benchmarks mainly assess LVLMs from a perception perspective, overlooking hallucinations arising from advanced reasoning capabilities. We develop the Perception-Reasoning Evaluation Hallucination (PRE-HAL) dataset, which enables the systematic evaluation of both perception and reasoning capabilities of LVLMs across multiple visual semantics, such as instances, scenes, and relations. Comprehensive evaluation with this new benchmark exposed more visual vulnerabilities, particularly in the more challenging task of relation reasoning. To address this issue, we propose, to the best of our knowledge, the first Dempster-Shafer theory (DST)-based visual hallucination detection method for LVLMs through uncertainty estimation. This method aims to efficiently capture the degree of conflict in high-level features at the model inference phase. Specifically, our approach employs simple mass functions to mitigate the computational complexity of evidence combination on power sets. We conduct an extensive evaluation of state-of-the-art LVLMs, LLaVA-v1.5, mPLUG-Owl2 and mPLUG-Owl3, with the new PRE-HAL benchmark. Experimental results indicate that our method outperforms five baseline uncertainty metrics, achieving average AUROC improvements of 4%, 10%, and 7% across three LVLMs. Our code is available at https://github.com/HT86159/Evidential-Conflict.
The representation gap between teacher and student is an emerging topic in knowledge distillation (KD). To reduce the gap and improve the performance, current methods often resort to complicated training schemes, loss functions, and feature alignments, which are task-specific and feature-specific. In this paper, we state that the essence of these methods is to discard the noisy information and distill the valuable information in the feature, and propose a novel KD method dubbed DiffKD, to explicitly denoise and match features using diffusion models. Our approach is based on the observation that student features typically contain more noises than teacher features due to the smaller capacity of student model. To address this, we propose to denoise student features using a diffusion model trained by teacher features. This allows us to perform better distillation between the refined clean feature and teacher feature. Additionally, we introduce a light-weight diffusion model with a linear autoencoder to reduce the computation cost and an adaptive noise matching module to improve the denoising performance. Extensive experiments demonstrate that DiffKD is effective across various types of features and achieves state-of-the-art performance consistently on image classification, object detection, and semantic segmentation tasks. Code is available at https://github.com/hunto/DiffKD.
Knowledge distillation is an effective paradigm for boosting the performance of pocket-size model, especially when multiple teacher models are available, the student would break the upper limit again. However, it is not economical to train diverse teacher models for the disposable distillation. In this paper, we introduce a new concept dubbed Avatars for distillation, which are the inference ensemble models derived from the teacher. Concretely, (1) For each iteration of distillation training, various Avatars are generated by a perturbation transformation. We validate that Avatars own higher upper limit of working capacity and teaching ability, aiding the student model in learning diverse and receptive knowledge perspectives from the teacher model. (2) During the distillation, we propose an uncertainty-aware factor from the variance of statistical differences between the vanilla teacher and Avatars, to adjust Avatars' contribution on knowledge transfer adaptively. Avatar Knowledge Distillation AKD is fundamentally different from existing methods and refines with the innovative view of unequal training. Comprehensive experiments demonstrate the effectiveness of our Avatars mechanism, which polishes up the state-of-the-art distillation methods for dense prediction without more extra computational cost. The AKD brings at most 0.7 AP gains on COCO 2017 for Object Detection and 1.83 mIoU gains on Cityscapes for Semantic Segmentation, respectively. Code is available at https://github.com/Gumpest/AvatarKD.
Evaluation metrics in machine learning are often hardly taken as loss functions, as they could be non-differentiable and non-decomposable, e.g., average precision and F1 score. This paper aims to address this problem by revisiting the surrogate loss learning, where a deep neural network is employed to approximate the evaluation metrics. Instead of pursuing an exact recovery of the evaluation metric through a deep neural network, we are reminded of the purpose of the existence of these evaluation metrics, which is to distinguish whether one model is better or worse than another. In this paper, we show that directly maintaining the relation of models between surrogate losses and metrics suffices, and propose a rank correlation-based optimization method to maximize this relation and learn surrogate losses. Compared to previous works, our method is much easier to optimize and enjoys significant efficiency and performance gains. Extensive experiments show that our method achieves improvements on various tasks including image classification and neural machine translation, and even outperforms state-of-the-art methods on human pose estimation and machine reading comprehension tasks. Code is available at: https://github.com/hunto/ReLoss.
One-shot neural architecture search (NAS) methods significantly reduce the search cost by considering the whole search space as one network, which only needs to be trained once. However, current methods select each operation independently without considering previous layers. Besides, the historical information obtained with huge computation cost is usually used only once and then discarded. In this paper, we introduce a sampling strategy based on Monte Carlo tree search (MCTS) with the search space modeled as a Monte Carlo tree (MCT), which captures the dependency among layers. Furthermore, intermediate results are stored in the MCT for the future decision and a better exploration-exploitation balance. Concretely, MCT is updated using the training loss as a reward to the architecture performance; for accurately evaluating the numerous nodes, we propose node communication and hierarchical node selection methods in the training and search stages, respectively, which make better uses of the operation rewards and hierarchical information. Moreover, for a fair comparison of different NAS methods, we construct an open-source NAS benchmark of a macro search space evaluated on CIFAR-10, namely NAS-Bench-Macro. Extensive experiments on NAS-Bench-Macro and ImageNet demonstrate that our method significantly improves search efficiency and performance. For example, by only searching $20$ architectures, our obtained architecture achieves $78.0\%$ top-1 accuracy with 442M FLOPs on ImageNet. Code (Benchmark) is available at: \url{https://github.com/xiusu/NAS-Bench-Macro}.
In coded aperture snapshot spectral imaging (CASSI) system, the real-world hyperspectral image (HSI) can be reconstructed from the captured compressive image in a snapshot. Model-based HSI reconstruction methods employed hand-crafted priors to solve the reconstruction problem, but most of which achieved limited success due to the poor representation capability of these hand-crafted priors. Deep learning based methods learning the mappings between the compressive images and the HSIs directly achieved much better results. Yet, it is nontrivial to design a powerful deep network heuristically for achieving satisfied results. In this paper, we propose a novel HSI reconstruction method based on the Maximum a Posterior (MAP) estimation framework using learned Gaussian Scale Mixture (GSM) prior. Different from existing GSM models using hand-crafted scale priors (e.g., the Jeffrey's prior), we propose to learn the scale prior through a deep convolutional neural network (DCNN). Furthermore, we also propose to estimate the local means of the GSM models by the DCNN. All the parameters of the MAP estimation algorithm and the DCNN parameters are jointly optimized through end-to-end training. Extensive experimental results on both synthetic and real datasets demonstrate that the proposed method outperforms existing state-of-the-art methods. The code is available at https://see.xidian.edu.cn/faculty/wsdong/Projects/DGSM-SCI.htm.
Creating an appearance-preserving level-of-detail (LoD) representation for arbitrary 3D scenes is a challenging problem. The appearance of a scene is an intricate combination of both geometry and material models, and is further complicated by correlation due to the spatial configuration of scene elements. We present a novel volumetric representation for the aggregated appearance of complex scenes and an efficient pipeline for LoD generation and rendering. The core of our representation is the Aggregated Bidirectional Scattering Distribution Function (ABSDF) that summarizes the far-field appearance of all surfaces inside a voxel. We propose a closed-form factorization of the ABSDF that accounts for spatially varying and orientation-varying material parameters. We tackle the challenge of capturing the correlation existing locally within a voxel and globally across different parts of the scene. Our method faithfully reproduces appearance and achieves higher quality than existing scene filtering methods while being inherently efficient to render. The memory footprint and rendering cost of our representation are independent of the original scene complexity.
As large language models (LLMs) are increasingly deployed in high-stakes applications, robust uncertainty estimation is essential for ensuring the safe and trustworthy deployment of LLMs. We present the most comprehensive study to date of uncertainty estimation in LLMs, evaluating 80 models spanning open- and closed-source families, dense and Mixture-of-Experts (MoE) architectures, reasoning and non-reasoning modes, quantization variants and parameter scales from 0.6B to 671B. Focusing on three representative black-box single-pass methods, including token probability-based uncertainty (TPU), numerical verbal uncertainty (NVU), and linguistic verbal uncertainty (LVU), we systematically evaluate uncertainty calibration and selective classification using the challenging MMLU-Pro benchmark, which covers both reasoning-intensive and knowledge-based tasks. Our results show that LVU consistently outperforms TPU and NVU, offering stronger calibration and discrimination while being more interpretable. We also find that high accuracy does not imply reliable uncertainty, and that model scale, post-training, reasoning ability and quantization all influence estimation performance. Notably, LLMs exhibit better uncertainty estimates on reasoning tasks than on knowledge-heavy ones, and good calibration does not necessarily translate to effective error ranking. These findings highlight the need for multi-perspective evaluation and position LVU as a practical tool for improving the reliability of LLMs in real-world settings.
Recently, zero-shot multi-label classification has garnered considerable attention for its capacity to operate predictions on unseen labels without human annotations. Nevertheless, prevailing approaches often use seen classes as imperfect proxies for unseen ones, resulting in suboptimal performance. Drawing inspiration from the success of text-to-image generation models in producing realistic images, we propose an innovative solution: generating synthetic data to construct a training set explicitly tailored for proxyless training on unseen labels. Our approach introduces a novel image generation framework that produces multi-label synthetic images of unseen classes for classifier training. To enhance diversity in the generated images, we leverage a pre-trained large language model to generate diverse prompts. Employing a pre-trained multi-modal CLIP model as a discriminator, we assess whether the generated images accurately represent the target classes. This enables automatic filtering of inaccurately generated images, preserving classifier accuracy. To refine text prompts for more precise and effective multi-label object generation, we introduce a CLIP score-based discriminative loss to fine-tune the text encoder in the diffusion model. Additionally, to enhance visual features on the target task while maintaining the generalization of original features and mitigating catastrophic forgetting resulting from fine-tuning the entire visual encoder, we propose a feature fusion module inspired by transformer attention mechanisms. This module aids in capturing global dependencies between multiple objects more effectively. Extensive experimental results validate the effectiveness of our approach, demonstrating significant improvements over state-of-the-art methods.
Masked autoencoders (MAEs) represent a prominent self-supervised learning paradigm in computer vision. Despite their empirical success, the underlying mechanisms of MAEs remain insufficiently understood. Recent studies have attempted to elucidate the functioning of MAEs through contrastive learning and feature representation analysis, yet these approaches often provide only implicit insights. In this paper, we propose a new perspective for understanding MAEs by leveraging the information bottleneck principle in information theory. Our theoretical analyses reveal that optimizing the latent features to balance relevant and irrelevant information is key to improving MAE performance. Building upon our proofs, we introduce MI-MAE, a novel method that optimizes MAEs through mutual information maximization and minimization. By enhancing latent features to retain maximal relevant information between them and the output, and minimizing irrelevant information between them and the input, our approach achieves better performance. Extensive experiments on standard benchmarks show that MI-MAE significantly outperforms MAE models in tasks such as image classification, object detection, and semantic segmentation. Our findings validate the theoretical framework and highlight the practical advantages of applying the information bottleneck principle to MAEs, offering deeper insights for developing more powerful self-supervised learning models.
Millimeter wave (mmWave) radar is a non-intrusive privacy and relatively convenient and inexpensive device, which has been demonstrated to be applicable in place of RGB cameras in human indoor pose estimation tasks. However, mmWave radar relies on the collection of reflected signals from the target, and the radar signals containing information is difficult to be fully applied. This has been a long-standing hindrance to the improvement of pose estimation accuracy. To address this major challenge, this paper introduces a probability map guided multi-format feature fusion model, ProbRadarM3F. This is a novel radar feature extraction framework using a traditional FFT method in parallel with a probability map based positional encoding method. ProbRadarM3F fuses the traditional heatmap features and the positional features, then effectively achieves the estimation of 14 keypoints of the human body. Experimental evaluation on the HuPR dataset proves the effectiveness of the model proposed in this paper, outperforming other methods experimented on this dataset with an AP of 69.9 %. The emphasis of our study is focusing on the position information that is not exploited before in radar singal. This provides direction to investigate other potential non-redundant information from mmWave rader.
Inspired by text prompts in large language models (LLMs), visual prompts have been explored to enhance the reasoning capabilities of large vision-language models (LVLMs). Current methods design heuristic visual prompts, such as overlaying a text-query-guided attention heatmap on the original input image. However, designing effective prompts manually is challenging and time-consuming, and it often fails to explore the benefits of different visual prompts, leading to sub-optimal performance. To this end, we propose \textbf{AutoV} that learns to automatically select the optimal visual prompt from various candidates based on given textual queries and the input image. To train AutoV, we developed an automatic data collection and labeling pipeline that evaluates various visual prompts with a pre-trained LVLM. We input a set of visual prompts into the LVLM and rank them according to the prediction losses generated by the model. Using the ranking as a supervision signal, we train AutoV to automatically choose the optimal visual prompt from various visual prompts for LVLMs. Experimental results indicate that AutoV enhances the performance of various LVLMs across multiple popular image understanding tasks. For instance, LLaVA-OV with AutoV achieves $\textbf{1.7}\%$ accuracy gain on LLaVA$^{\text{Wild}}$, and AutoV boosts Qwen2.5-VL by $\textbf{1.9}\%$ on MMMU, highlighting its potential as an optimal visual prompting method for LVLMs.
Remote sensing images (RSIs) capture both natural and human-induced changes on the Earth's surface, serving as essential data for environmental monitoring, urban planning, and resource management. Semantic segmentation (SS) of RSIs enables the fine-grained interpretation of surface features, making it a critical task in remote sensing analysis. With the increasing diversity and volume of RSIs collected by sensors on various platforms, traditional processing methods struggle to maintain efficiency and accuracy. In response, deep learning (DL) has emerged as a transformative approach, enabling substantial advances in remote sensing image semantic segmentation (RSISS) by automating feature extraction and improving segmentation accuracy across diverse modalities. This paper revisits the evolution of DL-based RSISS by categorizing existing approaches into four stages: the early pixel-based methods, the prevailing patch-based and tile-based techniques, and the emerging image-based strategies enabled by foundation models. We analyze these developments from the perspective of feature extraction and learning strategies, revealing the field's progression from pixel-level to tile-level and from unimodal to multimodal segmentation. Furthermore, we conduct a comprehensive evaluation of nearly 40 advanced techniques on a unified dataset to quantitatively characterize their performance and applicability. This review offers a holistic view of DL-based SS for RS, highlighting key advancements, comparative insights, and open challenges to guide future research.
Chenglizhao Chen, Shaofeng Liang, Runwei Guan et al.
Referring Multi-Object Tracking (RMOT) aims to achieve precise object detection and tracking through natural language instructions, representing a fundamental capability for intelligent robotic systems. However, current RMOT research remains mostly confined to ground-level scenarios, which constrains their ability to capture broad-scale scene contexts and perform comprehensive tracking and path planning. In contrast, Unmanned Aerial Vehicles (UAVs) leverage their expansive aerial perspectives and superior maneuverability to enable wide-area surveillance. Moreover, UAVs have emerged as critical platforms for Embodied Intelligence, which has given rise to an unprecedented demand for intelligent aerial systems capable of natural language interaction. To this end, we introduce AerialMind, the first large-scale RMOT benchmark in UAV scenarios, which aims to bridge this research gap. To facilitate its construction, we develop an innovative semi-automated collaborative agent-based labeling assistant (COALA) framework that significantly reduces labor costs while maintaining annotation quality. Furthermore, we propose HawkEyeTrack (HETrack), a novel method that collaboratively enhances vision-language representation learning and improves the perception of UAV scenarios. Comprehensive experiments validated the challenging nature of our dataset and the effectiveness of our method.
Existing optical flow datasets focus primarily on real-world simulation or synthetic human motion, but few are tailored to Celluloid(cel) anime character motion: a domain with unique visual and motion characteristics. To bridge this gap and facilitate research in optical flow estimation and downstream tasks such as anime video generation and line drawing colorization, we introduce LinkTo-Anime, the first high-quality dataset specifically designed for cel anime character motion generated with 3D model rendering. LinkTo-Anime provides rich annotations including forward and backward optical flow, occlusion masks, and Mixamo Skeleton. The dataset comprises 395 video sequences, totally 24,230 training frames, 720 validation frames, and 4,320 test frames. Furthermore, a comprehensive benchmark is constructed with various optical flow estimation methods to analyze the shortcomings and limitations across multiple datasets.
Accurate 3D semantic occupancy perception is essential for autonomous driving in complex environments with diverse and irregular objects. While vision-centric methods suffer from geometric inaccuracies, LiDAR-based approaches often lack rich semantic information. To address these limitations, MS-Occ, a novel multi-stage LiDAR-camera fusion framework which includes middle-stage fusion and late-stage fusion, is proposed, integrating LiDAR's geometric fidelity with camera-based semantic richness via hierarchical cross-modal fusion. The framework introduces innovations at two critical stages: (1) In the middle-stage feature fusion, the Gaussian-Geo module leverages Gaussian kernel rendering on sparse LiDAR depth maps to enhance 2D image features with dense geometric priors, and the Semantic-Aware module enriches LiDAR voxels with semantic context via deformable cross-attention; (2) In the late-stage voxel fusion, the Adaptive Fusion (AF) module dynamically balances voxel features across modalities, while the High Classification Confidence Voxel Fusion (HCCVF) module resolves semantic inconsistencies using self-attention-based refinement. Experiments on two large-scale benchmarks demonstrate state-of-the-art performance. On nuScenes-OpenOccupancy, MS-Occ achieves an Intersection over Union (IoU) of 32.1% and a mean IoU (mIoU) of 25.3%, surpassing the state-of-the-art by +0.7% IoU and +2.4% mIoU. Furthermore, on the SemanticKITTI benchmark, our method achieves a new state-of-the-art mIoU of 24.08%, robustly validating its generalization capabilities.Ablation studies further confirm the effectiveness of each individual module, highlighting substantial improvements in the perception of small objects and reinforcing the practical value of MS-Occ for safety-critical autonomous driving scenarios.
Accurate 3D multi-object tracking (MOT) is crucial for autonomous driving, as it enables robust perception, navigation, and planning in complex environments. While deep learning-based solutions have demonstrated impressive 3D MOT performance, model-based approaches remain appealing for their simplicity, interpretability, and data efficiency. Conventional model-based trackers typically rely on random vector-based Bayesian filters within the tracking-by-detection (TBD) framework but face limitations due to heuristic data association and track management schemes. In contrast, random finite set (RFS)-based Bayesian filtering handles object birth, survival, and death in a theoretically sound manner, facilitating interpretability and parameter tuning. In this paper, we present OptiPMB, a novel RFS-based 3D MOT method that employs an optimized Poisson multi-Bernoulli (PMB) filter while incorporating several key innovative designs within the TBD framework. Specifically, we propose a measurement-driven hybrid adaptive birth model for improved track initialization, employ adaptive detection probability parameters to effectively maintain tracks for occluded objects, and optimize density pruning and track extraction modules to further enhance overall tracking performance. Extensive evaluations on nuScenes and KITTI datasets show that OptiPMB achieves superior tracking accuracy compared with state-of-the-art methods, thereby establishing a new benchmark for model-based 3D MOT and offering valuable insights for future research on RFS-based trackers in autonomous driving.
Robust density estimation refers to the consistent estimation of the density function even when the data is contaminated by outliers. We find that existing forest density estimation at a certain point is inherently resistant to the outliers outside the cells containing the point, which we call \textit{non-local outliers}, but not resistant to the rest \textit{local outliers}. To achieve robustness against all outliers, we propose an ensemble learning algorithm called \textit{medians of forests for robust density estimation} (\textit{MFRDE}), which adopts a pointwise median operation on forest density estimators fitted on subsampled datasets. Compared to existing robust kernel-based methods, MFRDE enables us to choose larger subsampling sizes, sacrificing less accuracy for density estimation while achieving robustness. On the theoretical side, we introduce the local outlier exponent to quantify the number of local outliers. Under this exponent, we show that even if the number of outliers reaches a certain polynomial order in the sample size, MFRDE is able to achieve almost the same convergence rate as the same algorithm on uncontaminated data, whereas robust kernel-based methods fail. On the practical side, real data experiments show that MFRDE outperforms existing robust kernel-based methods. Moreover, we apply MFRDE to anomaly detection to showcase a further application.
Learning informative representations from image-based observations is of fundamental concern in deep Reinforcement Learning (RL). However, data-inefficiency remains a significant barrier to this objective. To overcome this obstacle, we propose to accelerate state representation learning by enforcing view-consistency on the dynamics. Firstly, we introduce a formalism of Multi-view Markov Decision Process (MMDP) that incorporates multiple views of the state. Following the structure of MMDP, our method, View-Consistent Dynamics (VCD), learns state representations by training a view-consistent dynamics model in the latent space, where views are generated by applying data augmentation to states. Empirical evaluation on DeepMind Control Suite and Atari-100k demonstrates VCD to be the SoTA data-efficient algorithm on visual control tasks.
Training a good supernet in one-shot NAS methods is difficult since the search space is usually considerably huge (e.g., $13^{21}$). In order to enhance the supernet's evaluation ability, one greedy strategy is to sample good paths, and let the supernet lean towards the good ones and ease its evaluation burden as a result. However, in practice the search can be still quite inefficient since the identification of good paths is not accurate enough and sampled paths still scatter around the whole search space. In this paper, we leverage an explicit path filter to capture the characteristics of paths and directly filter those weak ones, so that the search can be thus implemented on the shrunk space more greedily and efficiently. Concretely, based on the fact that good paths are much less than the weak ones in the space, we argue that the label of "weak paths" will be more confident and reliable than that of "good paths" in multi-path sampling. In this way, we thus cast the training of path filter in the positive and unlabeled (PU) learning paradigm, and also encourage a \textit{path embedding} as better path/operation representation to enhance the identification capacity of the learned filter. By dint of this embedding, we can further shrink the search space by aggregating similar operations with similar embeddings, and the search can be more efficient and accurate. Extensive experiments validate the effectiveness of the proposed method GreedyNASv2. For example, our obtained GreedyNASv2-L achieves $81.1\%$ Top-1 accuracy on ImageNet dataset, significantly outperforming the ResNet-50 strong baselines.
Automotive radar provides reliable environmental perception in all-weather conditions with affordable cost, but it hardly supplies semantic and geometry information due to the sparsity of radar detection points. With the development of automotive radar technologies in recent years, instance segmentation becomes possible by using automotive radar. Its data contain contexts such as radar cross section and micro-Doppler effects, and sometimes can provide detection when the field of view is obscured. The outcome from instance segmentation could be potentially used as the input of trackers for tracking targets. The existing methods often utilize a clustering-based classification framework, which fits the need of real-time processing but has limited performance due to minimum information provided by sparse radar detection points. In this paper, we propose an efficient method based on clustering of estimated semantic information to achieve instance segmentation for the sparse radar detection points. In addition, we show that the performance of the proposed approach can be further enhanced by incorporating the visual multi-layer perceptron. The effectiveness of the proposed method is verified by experimental results on the popular RadarScenes dataset, achieving 89.53% mean coverage and 86.97% mean average precision with the IoU threshold of 0.5, which is superior to other approaches in the literature. More significantly, the consumed memory is around 1MB, and the inference time is less than 40ms, indicating that our proposed algorithm is storage and time efficient. These two criteria ensure the practicality of the proposed method in real-world systems.
In this paper, we propose a gradient boosting algorithm for large-scale regression problems called \textit{Gradient Boosted Binary Histogram Ensemble} (GBBHE) based on binary histogram partition and ensemble learning. From the theoretical perspective, by assuming the Hölder continuity of the target function, we establish the statistical convergence rate of GBBHE in the space $C^{0,α}$ and $C^{1,0}$, where a lower bound of the convergence rate for the base learner demonstrates the advantage of boosting. Moreover, in the space $C^{1,0}$, we prove that the number of iterations to achieve the fast convergence rate can be reduced by using ensemble regressor as the base learner, which improves the computational efficiency. In the experiments, compared with other state-of-the-art algorithms such as gradient boosted regression tree (GBRT), Breiman's forest, and kernel-based methods, our GBBHE algorithm shows promising performance with less running time on large-scale datasets.
Searching for network width is an effective way to slim deep neural networks with hardware budgets. With this aim, a one-shot supernet is usually leveraged as a performance evaluator to rank the performance \wrt~different width. Nevertheless, current methods mainly follow a manually fixed weight sharing pattern, which is limited to distinguish the performance gap of different width. In this paper, to better evaluate each width, we propose a locally free weight sharing strategy (CafeNet) accordingly. In CafeNet, weights are more freely shared, and each width is jointly indicated by its base channels and free channels, where free channels are supposed to loCAte FrEely in a local zone to better represent each width. Besides, we propose to further reduce the search space by leveraging our introduced FLOPs-sensitive bins. As a result, our CafeNet can be trained stochastically and get optimized within a min-min strategy. Extensive experiments on ImageNet, CIFAR-10, CelebA and MS COCO dataset have verified our superiority comparing to other state-of-the-art baselines. For example, our method can further boost the benchmark NAS network EfficientNet-B0 by 0.41\% via searching its width more delicately.
Differentiable neural architecture search (DARTS) has gained much success in discovering flexible and diverse cell types. To reduce the evaluation gap, the supernet is expected to have identical layers with the target network. However, even for this consistent search, the searched cells often suffer from poor performance, especially for the supernet with fewer layers, as current DARTS methods are prone to wide and shallow cells, and this topology collapse induces sub-optimal searched cells. In this paper, we alleviate this issue by endowing the cells with explicit stretchability, so the search can be directly implemented on our stretchable cells for both operation type and topology simultaneously. Concretely, we introduce a set of topological variables and a combinatorial probabilistic distribution to explicitly model the target topology. With more diverse and complex topologies, our method adapts well for various layer numbers. Extensive experiments on CIFAR-10 and ImageNet show that our stretchable cells obtain better performance with fewer layers and parameters. For example, our method can improve DARTS by 0.28\% accuracy on CIFAR-10 dataset with 45\% parameters reduced or 2.9\% with similar FLOPs on ImageNet dataset.
To deploy a well-trained CNN model on low-end computation edge devices, it is usually supposed to compress or prune the model under certain computation budget (e.g., FLOPs). Current filter pruning methods mainly leverage feature maps to generate important scores for filters and prune those with smaller scores, which ignores the variance of input batches to the difference in sparse structure over filters. In this paper, we propose a data agnostic filter pruning method that uses an auxiliary network named Dagger module to induce pruning and takes pretrained weights as input to learn the importance of each filter. In addition, to help prune filters with certain FLOPs constraints, we leverage an explicit FLOPs-aware regularization to directly promote pruning filters toward target FLOPs. Extensive experimental results on CIFAR-10 and ImageNet datasets indicate our superiority to other state-of-the-art filter pruning methods. For example, our 50\% FLOPs ResNet-50 can achieve 76.1\% Top-1 accuracy on ImageNet dataset, surpassing many other filter pruning methods.
Variational quantum algorithms (VQAs) are expected to be a path to quantum advantages on noisy intermediate-scale quantum devices. However, both empirical and theoretical results exhibit that the deployed ansatz heavily affects the performance of VQAs such that an ansatz with a larger number of quantum gates enables a stronger expressivity, while the accumulated noise may render a poor trainability. To maximally improve the robustness and trainability of VQAs, here we devise a resource and runtime efficient scheme termed quantum architecture search (QAS). In particular, given a learning task, QAS automatically seeks a near-optimal ansatz (i.e., circuit architecture) to balance benefits and side-effects brought by adding more noisy quantum gates to achieve a good performance. We implement QAS on both the numerical simulator and real quantum hardware, via the IBM cloud, to accomplish data classification and quantum chemistry tasks. In the problems studied, numerical and experimental results show that QAS can not only alleviate the influence of quantum noise and barren plateaus, but also outperforms VQAs with pre-selected ansatze.
Training a supernet matters for one-shot neural architecture search (NAS) methods since it serves as a basic performance estimator for different architectures (paths). Current methods mainly hold the assumption that a supernet should give a reasonable ranking over all paths. They thus treat all paths equally, and spare much effort to train paths. However, it is harsh for a single supernet to evaluate accurately on such a huge-scale search space (e.g., $7^{21}$). In this paper, instead of covering all paths, we ease the burden of supernet by encouraging it to focus more on evaluation of those potentially-good ones, which are identified using a surrogate portion of validation data. Concretely, during training, we propose a multi-path sampling strategy with rejection, and greedily filter the weak paths. The training efficiency is thus boosted since the training space has been greedily shrunk from all paths to those potentially-good ones. Moreover, we further adopt an exploration and exploitation policy by introducing an empirical candidate path pool. Our proposed method GreedyNAS is easy-to-follow, and experimental results on ImageNet dataset indicate that it can achieve better Top-1 accuracy under same search space and FLOPs or latency level, but with only $\sim$60\% of supernet training cost. By searching on a larger space, our GreedyNAS can also obtain new state-of-the-art architectures.