Shihao Zhang

Marwane Hariat, Olivier Laurent, Rémi Kazmierczak et al.

Semantic segmentation methods have advanced significantly. Still, their robustness to real-world perturbations and object types not seen during training remains a challenge, particularly in safety-critical applications. We propose a novel approach to improve the robustness of semantic segmentation techniques by leveraging the synergy between label-to-image generators and image-to-label segmentation models. Specifically, we design Robusta, a novel robust conditional generative adversarial network to generate realistic and plausible perturbed images that can be used to train reliable segmentation models. We conduct in-depth studies of the proposed generative model, assess the performance and robustness of the downstream segmentation network, and demonstrate that our approach can significantly enhance the robustness in the face of real-world perturbations, distribution shifts, and out-of-distribution samples. Our results suggest that this approach could be valuable in safety-critical applications, where the reliability of perception modules such as semantic segmentation is of utmost importance and comes with a limited computational budget in inference. We release our code at https://github.com/ENSTA-U2IS-AI/robusta.

Shihao Zhang, Rayan Saab

Post-training quantization is widely used for compressing large neural networks, but aggressive low-bit quantization can significantly degrade model quality. A common remedy is to augment the quantized weights with a low-rank correction, leading to approximations of the form $W\approx Q+LR$. In this paper, we study this low-precision plus low-rank representation through the layer-wise reconstruction objective $\|XW-X(Q+LR)\|_F^2$, where $X$ is a calibration matrix. We establish, to our knowledge, the first information-theoretic lower bounds for this problem under finite-alphabet and bounded low-rank compensation constraints. We then propose GPTQ-intrinsic LoRA, a training-free algorithm that incorporates the low-rank correction directly into a GPTQ-style quantization pass by appropriately augmenting the calibration Hessian. For the choice $L=V_r$, where $V_r$ contains the top right singular vectors of $X$, we prove layer-wise reconstruction error bounds in which the usual GPTQ dependence on $\|X\|_F^2$ is replaced by the rank-$r$ residual $\|X-X_r\|_F^2$, up to regularization terms. Under natural structural assumptions, these bounds match the information-theoretic lower bounds in their dominant scaling, up to constants and mild factors. We also introduce Bid-Up, a fixed-grid quantization refinement step that can be alternated with optimal low-rank compensation with guaranteed non-increasing layer-wise reconstruction error. Experiments on Qwen3 language models and DeiT vision transformers show that GPTQ-intrinsic LoRA improves over GPTQ and GPTQ followed by low-rank compensation, with additional gains from refinement loops.

Chen Minqi, Zhongqi Yue, Shihao Zhang et al.

Rotary Position Embedding (RoPE) has become a core component of modern Transformer architectures across language, vision, and 3D domains. However, existing implementations rely on vector-level split and merge operations that introduce non-negligible computational overhead, often overlooked in attention optimization. The problem is further amplified in multi-dimensional settings (e.g., 2D and 3D RoPE), where additional vector operations and uneven feature partitions degrade hardware utilization. To overcome these limitations, we propose RoME (Rotary Matrix position Embedding), a mathematically equivalent yet computationally efficient reformulation of RoPE that replaces vector operations with unified matrix transformations. RoME eliminates dimension-specific operations, simplifies implementation, and enables fused parallel execution across Cube and Vector units on modern NPUs. Experiments show that RoME delivers substantial acceleration at both the operator and full-model levels. The implementation is available at https://gitcode.com/cann/ops-transformer/blob/master/experimental/posembedding/rope_matrix/README.md.

Shihao Zhang, Linlin Yang, Michael Bi Mi et al.

In computer vision, it is often observed that formulating regression problems as a classification task often yields better performance. We investigate this curious phenomenon and provide a derivation to show that classification, with the cross-entropy loss, outperforms regression with a mean squared error loss in its ability to learn high-entropy feature representations. Based on the analysis, we propose an ordinal entropy loss to encourage higher-entropy feature spaces while maintaining ordinal relationships to improve the performance of regression tasks. Experiments on synthetic and real-world regression tasks demonstrate the importance and benefits of increasing entropy for regression.

Shihao Zhang, Ziwei Wang, Jie Zhou et al.

While Aspect-based Sentiment Analysis (ABSA) systems have achieved high accuracy in identifying sentiment polarities, they often operate as "black boxes," lacking the explicit reasoning capabilities characteristic of human affective cognition. Humans do not merely categorize sentiment; they construct causal explanations for their judgments. To bridge this gap, we propose ABSA-R1, a large language model framework designed to mimic this ``reason-before-predict" cognitive process. By leveraging reinforcement learning (RL), ABSA-R1 learns to articulate the why behind the what, generating natural language justifications that ground its sentiment predictions. We introduce a Cognition-Aligned Reward Model (formerly sentiment-aware reward model) that enforces consistency between the generated reasoning path and the final emotional label. Furthermore, inspired by metacognitive monitoring, we implement a performance-driven rejection sampling strategy that selectively targets hard cases where the model's internal reasoning is uncertain or inconsistent. Experimental results on four benchmarks demonstrate that equipping models with this explicit reasoning capability not only enhances interpretability but also yields superior performance in sentiment classification and triplet extraction compared to non-reasoning baselines.

Shihao Zhang, kenji kawaguchi, Angela Yao

Most works studying representation learning focus only on classification and neglect regression. Yet, the learning objectives and, therefore, the representation topologies of the two tasks are fundamentally different: classification targets class separation, leading to disconnected representations, whereas regression requires ordinality with respect to the target, leading to continuous representations. We thus wonder how the effectiveness of a regression representation is influenced by its topology, with evaluation based on the Information Bottleneck (IB) principle. The IB principle is an important framework that provides principles for learning effective representations. We establish two connections between it and the topology of regression representations. The first connection reveals that a lower intrinsic dimension of the feature space implies a reduced complexity of the representation Z. This complexity can be quantified as the conditional entropy of Z on the target Y, and serves as an upper bound on the generalization error. The second connection suggests a feature space that is topologically similar to the target space will better align with the IB principle. Based on these two connections, we introduce PH-Reg, a regularizer specific to regression that matches the intrinsic dimension and topology of the feature space with the target space. Experiments on synthetic and real-world regression tasks demonstrate the benefits of PH-Reg. Code: https://github.com/needylove/PH-Reg.

Shihao Zhang, Yuguang Yan, Angela Yao

For deep regression, preserving the ordinality of the targets with respect to the feature representation improves performance across various tasks. However, a theoretical explanation for the benefits of ordinality is still lacking. This work reveals that preserving ordinality reduces the conditional entropy $H(Z|Y)$ of representation $Z$ conditional on the target $Y$. However, our findings reveal that typical regression losses do little to reduce $H(Z|Y)$, even though it is vital for generalization performance. With this motivation, we introduce an optimal transport-based regularizer to preserve the similarity relationships of targets in the feature space to reduce $H(Z|Y)$. Additionally, we introduce a simple yet efficient strategy of duplicating the regressor targets, also with the aim of reducing $H(Z|Y)$. Experiments on three real-world regression tasks verify the effectiveness of our strategies to improve deep regression. Code: https://github.com/needylove/Regression_tightness.

Haoyu Zhang, Shihao Zhang, Ian Colbert et al.

Post-training quantization (PTQ) has become a crucial tool for reducing the memory and compute costs of modern deep neural networks, including large language models (LLMs). Among PTQ algorithms, the OPTQ framework-also known as GPTQ-has emerged as a leading method due to its computational efficiency and strong empirical performance. Despite its widespread adoption, however, OPTQ lacks rigorous quantitative theoretical guarantees. This paper presents the first quantitative error bounds for both deterministic and stochastic variants of OPTQ, as well as for Qronos, a recent related state-of-the-art PTQ algorithm. We analyze how OPTQ's iterative procedure induces quantization error and derive non-asymptotic 2-norm error bounds that depend explicitly on the calibration data and a regularization parameter that OPTQ uses. Our analysis provides theoretical justification for several practical design choices, including the widely used heuristic of ordering features by decreasing norm, as well as guidance for selecting the regularization parameter. For the stochastic variant, we establish stronger infinity-norm error bounds, which enable control over the required quantization alphabet and are particularly useful for downstream layers and nonlinearities. Finally, we extend our analysis to Qronos, providing new theoretical bounds, for both its deterministic and stochastic variants, that help explain its empirical advantages.

Shihao Zhang, Rayan Saab

Deep neural networks have achieved state-of-the-art performance across numerous applications, but their high memory and computational demands present significant challenges, particularly in resource-constrained environments. Model compression techniques, such as low-rank approximation, offer a promising solution by reducing the size and complexity of these networks while only minimally sacrificing accuracy. In this paper, we develop an analytical framework for data-driven post-training low-rank compression. We prove three recovery theorems under progressively weaker assumptions about the approximate low-rank structure of activations, modeling deviations via noise. Our results represent a step toward explaining why data-driven low-rank compression methods outperform data-agnostic approaches and towards theoretically grounded compression algorithms that reduce inference costs while maintaining performance.

Shuai Wang, Shihao Zhang, Jiaqi Wu et al.

Clear monitoring images are crucial for the safe operation of coal mine Internet of Video Things (IoVT) systems. However, low illumination and uneven brightness in underground environments significantly degrade image quality, posing challenges for enhancement methods that often rely on difficult-to-obtain paired reference images. Additionally, there is a trade-off between enhancement performance and computational efficiency on edge devices within IoVT systems.To address these issues, we propose a multimodal image enhancement method tailored for coal mine IoVT, utilizing an ISP-CNN fusion architecture optimized for uneven illumination. This two-stage strategy combines global enhancement with detail optimization, effectively improving image quality, especially in poorly lit areas. A CLIP-based multimodal iterative optimization allows for unsupervised training of the enhancement algorithm. By integrating traditional image signal processing (ISP) with convolutional neural networks (CNN), our approach reduces computational complexity while maintaining high performance, making it suitable for real-time deployment on edge devices.Experimental results demonstrate that our method effectively mitigates uneven brightness and enhances key image quality metrics, with PSNR improvements of 2.9%-4.9%, SSIM by 4.3%-11.4%, and VIF by 4.9%-17.8% compared to seven state-of-the-art algorithms. Simulated coal mine monitoring scenarios validate our method's ability to balance performance and computational demands, facilitating real-time enhancement and supporting safer mining operations.

Shihao Zhang, Rayan Saab

Quantization is a widely used compression technique for reducing the memory and computation costs of large pre-trained models. A key challenge in per-channel post-training quantization (PTQ) is selecting appropriate scaling factors to replace weight values with values from a scaled integer grid. Existing methods typically fix the scale at the outset via heuristic tuning or grid search. We propose Beacon, a simple and effective algorithm that eliminates the need for such manual tuning. Beacon performs per-channel PTQ directly using an unscaled grid and automatically determines the optimal scaling factors by exploiting the geometry of scalar quantization. It does not rely on back-propagation or large calibration sets. Despite its simplicity and tuning-free nature, Beacon achieves competitive performance compared to state-of-the-art methods, making it a practical solution for efficient model deployment.

Jiaqi Wu, Shihao Zhang, Simin Chen et al.

Edge computing has emerged as a key paradigm for deploying deep learning-based object detection in time-sensitive scenarios. However, existing edge detection methods face challenges: 1) difficulty balancing detection precision with lightweight models, 2) limited adaptability of generalized deployment designs, and 3) insufficient real-world validation. To address these issues, we propose the Edge Detection Toolbox (ED-TOOLBOX), which utilizes generalizable plug-and-play components to adapt object detection models for edge environments. Specifically, we introduce a lightweight Reparameterized Dynamic Convolutional Network (Rep-DConvNet) featuring weighted multi-shape convolutional branches to enhance detection performance. Additionally, we design a Sparse Cross-Attention (SC-A) network with a localized-mapping-assisted self-attention mechanism, enabling a well-crafted joint module for adaptive feature transfer. For real-world applications, we incorporate an Efficient Head into the YOLO framework to accelerate edge model optimization. To demonstrate practical impact, we identify a gap in helmet detection -- overlooking band fastening, a critical safety factor -- and create the Helmet Band Detection Dataset (HBDD). Using ED-TOOLBOX-optimized models, we address this real-world task. Extensive experiments validate the effectiveness of ED-TOOLBOX, with edge detection models outperforming six state-of-the-art methods in visual surveillance simulations, achieving real-time and accurate performance. These results highlight ED-TOOLBOX as a superior solution for edge object detection.

Bharath Ramesh, Shihao Zhang, Hong Yang et al.

This paper presents a long-term object tracking framework with a moving event camera under general tracking conditions. A first of its kind for these revolutionary cameras, the tracking framework uses a discriminative representation for the object with online learning, and detects and re-tracks the object when it comes back into the field-of-view. One of the key novelties is the use of an event-based local sliding window technique that tracks reliably in scenes with cluttered and textured background. In addition, Bayesian bootstrapping is used to assist real-time processing and boost the discriminative power of the object representation. On the other hand, when the object re-enters the field-of-view of the camera, a data-driven, global sliding window detector locates the object for subsequent tracking. Extensive experiments demonstrate the ability of the proposed framework to track and detect arbitrary objects of various shapes and sizes, including dynamic objects such as a human. This is a significant improvement compared to earlier works that simply track objects as long as they are visible under simpler background settings. Using the ground truth locations for five different objects under three motion settings, namely translation, rotation and 6-DOF, quantitative measurement is reported for the event-based tracking framework with critical insights on various performance issues. Finally, real-time implementation in C++ highlights tracking ability under scale, rotation, view-point and occlusion scenarios in a lab setting.

Shihao Zhang, Huazhu Fu, Yanwu Xu et al.

Retinal image segmentation plays an important role in automatic disease diagnosis. This task is very challenging because the complex structure and texture information are mixed in a retinal image, and distinguishing the information is difficult. Existing methods handle texture and structure jointly, which may lead biased models toward recognizing textures and thus results in inferior segmentation performance. To address it, we propose a segmentation strategy that seeks to separate structure and texture components and significantly improve the performance. To this end, we design a structure-texture demixing network (STD-Net) that can process structures and textures differently and better. Extensive experiments on two retinal image segmentation tasks (i.e., blood vessel segmentation, optic disc and cup segmentation) demonstrate the effectiveness of the proposed method.

Huazhu Fu, Fei Li, Xu Sun et al.

Angle closure glaucoma (ACG) is a more aggressive disease than open-angle glaucoma, where the abnormal anatomical structures of the anterior chamber angle (ACA) may cause an elevated intraocular pressure and gradually lead to glaucomatous optic neuropathy and eventually to visual impairment and blindness. Anterior Segment Optical Coherence Tomography (AS-OCT) imaging provides a fast and contactless way to discriminate angle closure from open angle. Although many medical image analysis algorithms have been developed for glaucoma diagnosis, only a few studies have focused on AS-OCT imaging. In particular, there is no public AS-OCT dataset available for evaluating the existing methods in a uniform way, which limits progress in the development of automated techniques for angle closure detection and assessment. To address this, we organized the Angle closure Glaucoma Evaluation challenge (AGE), held in conjunction with MICCAI 2019. The AGE challenge consisted of two tasks: scleral spur localization and angle closure classification. For this challenge, we released a large dataset of 4800 annotated AS-OCT images from 199 patients, and also proposed an evaluation framework to benchmark and compare different models. During the AGE challenge, over 200 teams registered online, and more than 1100 results were submitted for online evaluation. Finally, eight teams participated in the onsite challenge. In this paper, we summarize these eight onsite challenge methods and analyze their corresponding results for the two tasks. We further discuss limitations and future directions. In the AGE challenge, the top-performing approach had an average Euclidean Distance of 10 pixels (10um) in scleral spur localization, while in the task of angle closure classification, all the algorithms achieved satisfactory performances, with two best obtaining an accuracy rate of 100%.

Shihao Zhang, Huazhu Fu, Yuguang Yan et al.

Learning structural information is critical for producing an ideal result in retinal image segmentation. Recently, convolutional neural networks have shown a powerful ability to extract effective representations. However, convolutional and pooling operations filter out some useful structural information. In this paper, we propose an Attention Guided Network (AG-Net) to preserve the structural information and guide the expanding operation. In our AG-Net, the guided filter is exploited as a structure sensitive expanding path to transfer structural information from previous feature maps, and an attention block is introduced to exclude the noise and reduce the negative influence of background further. The extensive experiments on two retinal image segmentation tasks (i.e., blood vessel segmentation, optic disc and cup segmentation) demonstrate the effectiveness of our proposed method.

Bharath Ramesh, Hong Yang, Garrick Orchard et al.

We introduce a generic visual descriptor, termed as distribution aware retinal transform (DART), that encodes the structural context using log-polar grids for event cameras. The DART descriptor is applied to four different problems, namely object classification, tracking, detection and feature matching: (1) The DART features are directly employed as local descriptors in a bag-of-features classification framework and testing is carried out on four standard event-based object datasets (N-MNIST, MNIST-DVS, CIFAR10-DVS, NCaltech-101). (2) Extending the classification system, tracking is demonstrated using two key novelties: (i) For overcoming the low-sample problem for the one-shot learning of a binary classifier, statistical bootstrapping is leveraged with online learning; (ii) To achieve tracker robustness, the scale and rotation equivariance property of the DART descriptors is exploited for the one-shot learning. (3) To solve the long-term object tracking problem, an object detector is designed using the principle of cluster majority voting. The detection scheme is then combined with the tracker to result in a high intersection-over-union score with augmented ground truth annotations on the publicly available event camera dataset. (4) Finally, the event context encoded by DART greatly simplifies the feature correspondence problem, especially for spatio-temporal slices far apart in time, which has not been explicitly tackled in the event-based vision domain.

Shihao Zhang, Weiyao Lin, Ping Lu et al.

Object detection is an important yet challenging task in video understanding & analysis, where one major challenge lies in the proper balance between two contradictive factors: detection accuracy and detection speed. In this paper, we propose a new adaptive patch-of-interest composition approach for boosting both the accuracy and speed for object detection. The proposed approach first extracts patches in a video frame which have the potential to include objects-of-interest. Then, an adaptive composition process is introduced to compose the extracted patches into an optimal number of sub-frames for object detection. With this process, we are able to maintain the resolution of the original frame during object detection (for guaranteeing the accuracy), while minimizing the number of inputs in detection (for boosting the speed). Experimental results on various datasets demonstrate the effectiveness of the proposed approach.