Hanyang Peng

Wanli Zhong, Haibo Feng, Zirui Zhou et al.

Deploying Transformer models on edge devices is limited by latency and energy budgets. While INT8 quantization effectively accelerates the primary matrix multiplications, it exposes the softmax as the dominant bottleneck. This stage incurs a costly dequantize-softmax-requantize detour, which can account for up to 65% of total attention latency and disrupts the end-to-end integer dataflow critical for edge hardware efficiency. To address this limitation, we present IntAttention, the first fully integer, plug-and-play attention pipeline without retraining. At the core of our approach lies IndexSoftmax, a hardware-friendly operator that replaces floating-point exponentials entirely within the integer domain. IntAttention integrates sparsity-aware clipping, a 32-entry lookup-table approximation, and direct integer normalization, thereby eliminating all datatype conversion overhead. We evaluate IntAttention and demonstrate consistent and substantial gains. Our method achieves up to 3.7x speedup and 61% energy reduction over FP16 baselines and 2.0x faster than conventional INT8 attention pipelines on Armv8 CPUs. These gains are achieved with high-fidelity accuracy comparable to baselines across diverse language and vision models, enabling practical and efficient Transformer inference on commodity edge devices. Code will be released in later version of this work.

Haiyun Li, Zhiyong Wu, Xiaofeng Xie et al.

Voice cloning (VC)-resistant watermarking is an emerging technique for tracing and preventing unauthorized cloning. Existing methods effectively trace traditional VC models by training them on watermarked audio but fail in zero-shot VC scenarios, where models synthesize audio from an audio prompt without training. To address this, we propose VoiceMark, the first zero-shot VC-resistant watermarking method that leverages speaker-specific latents as the watermark carrier, allowing the watermark to transfer through the zero-shot VC process into the synthesized audio. Additionally, we introduce VC-simulated augmentations and VAD-based loss to enhance robustness against distortions. Experiments on multiple zero-shot VC models demonstrate that VoiceMark achieves over 95% accuracy in watermark detection after zero-shot VC synthesis, significantly outperforming existing methods, which only reach around 50%. See our code and demos at: https://huggingface.co/spaces/haiyunli/VoiceMark

Jiesi Hu, Yanwu Yang, Zhiyu Ye et al.

Universal models for medical image segmentation, such as interactive and in-context learning (ICL) models, offer strong generalization but require extensive annotations. Interactive models need repeated user prompts for each image, while ICL relies on dense, pixel-level labels. To address this, we propose Weakly Supervised In-Context Learning (WS-ICL), a new ICL paradigm that leverages weak prompts (e.g., bounding boxes or points) instead of dense labels for context. This approach significantly reduces annotation effort by eliminating the need for fine-grained masks and repeated user prompting for all images. We evaluated the proposed WS-ICL model on three held-out benchmarks. Experimental results demonstrate that WS-ICL achieves performance comparable to regular ICL models at a significantly lower annotation cost. In addition, WS-ICL is highly competitive even under the interactive paradigm. These findings establish WS-ICL as a promising step toward more efficient and unified universal models for medical image segmentation. Our code and model are publicly available at https://github.com/jiesihu/Weak-ICL.

Jiesi Hu, Chenfei Ye, Yanwu Yang et al.

In-context learning (ICL), a type of universal model, demonstrates exceptional generalization across a wide range of tasks without retraining by leveraging task-specific guidance from context, making it particularly effective for the intricate demands of neuroimaging. However, current ICL models, limited to 2D inputs and thus exhibiting suboptimal performance, struggle to extend to 3D inputs due to the high memory demands of ICL. In this regard, we introduce Neuroverse3D, an ICL model capable of performing multiple neuroimaging tasks in 3D (e.g., segmentation, denoising, inpainting). Neuroverse3D overcomes the large memory consumption associated with 3D inputs through adaptive parallel-sequential context processing and a U-shaped fusion strategy, allowing it to handle an unlimited number of context images. Additionally, we propose an optimized loss function to balance multi-task training and enhance focus on anatomical boundaries. Our study incorporates 43,674 3D multi-modal scans from 19 neuroimaging datasets and evaluates Neuroverse3D on 14 diverse tasks using held-out test sets. The results demonstrate that Neuroverse3D significantly outperforms existing ICL models and closely matches task-specific models, enabling flexible adaptation to medical center variations without retraining. The code and model weights are publicly available at https://github.com/jiesihu/Neuroverse3D.

Hanbo Bi, Yingchao Feng, Boyuan Tong et al.

The rapid advancement of foundation models has revolutionized visual representation learning in a self-supervised manner. However, their application in remote sensing (RS) remains constrained by a fundamental gap: existing models predominantly handle single or limited modalities, overlooking the inherently multi-modal nature of RS observations. Optical, synthetic aperture radar (SAR), and multi-spectral data offer complementary insights that significantly reduce the inherent ambiguity and uncertainty in single-source analysis. To bridge this gap, we introduce RingMoE, a unified multi-modal RS foundation model with 14.7 billion parameters, pre-trained on 400 million multi-modal RS images from nine satellites. RingMoE incorporates three key innovations: (1) A hierarchical Mixture-of-Experts (MoE) architecture comprising modal-specialized, collaborative, and shared experts, effectively modeling intra-modal knowledge while capturing cross-modal dependencies to mitigate conflicts between modal representations; (2) Physics-informed self-supervised learning, explicitly embedding sensor-specific radiometric characteristics into the pre-training objectives; (3) Dynamic expert pruning, enabling adaptive model compression from 14.7B to 1B parameters while maintaining performance, facilitating efficient deployment in Earth observation applications. Evaluated across 23 benchmarks spanning six key RS tasks (i.e., classification, detection, segmentation, tracking, change detection, and depth estimation), RingMoE outperforms existing foundation models and sets new SOTAs, demonstrating remarkable adaptability from single-modal to multi-modal scenarios. Beyond theoretical progress, it has been deployed and trialed in multiple sectors, including emergency response, land management, marine sciences, and urban planning.

Hanyang Peng, Shuang Qin, Yue Yu et al.

Adam has proven remarkable successful in training deep neural networks, but the mechanisms underlying its empirical successes and limitations remain underexplored. In this study, we demonstrate that the effectiveness of Adam stems largely from its similarity to SignSGD in robustly handling large gradient fluctuations, yet it is also vulnerable to destabilizing loss spikes due to its uncontrolled update scaling. To enhance the advantage of Adam and mitigate its limitation, we propose SignSoftSGD (S3), a novel optimizer with three key innovations. \emph{First}, S3 generalizes the sign-like update by employing a flexible $p$-th order momentum ($p \geq 1$) in the denominator, departing from the conventional second-order momentum (variance) preconditioning. This design enables enhanced performance while achieving stable training even with aggressive learning rates. \emph{Second}, S3 minimizes the occurrences of loss spikes through unified exponential moving average coefficients for numerator and denominator momenta, which inherently bound updates to $[-1, 1]$ and simplify hyperparameter tuning. \emph{Third}, S3 incorporates an equivalent Nesterov's accelerated gradient(NAG) module, accelerating convergence without memory overhead. Theoretically, we prove that S3 achieves the optimal convergence rate of $O\left(\frac{1}{T^{\sfrac{1}{4}}}\right)$ for general nonconvex stochastic optimization under weak assumptions. Extensive experiments across a range of vision and language tasks show that \textsf{\small S3} not only converges more rapidly and improves performance but also rarely experiences loss spikes, even with a \textbf{$\bm{10 \times}$} larger learning rate. In fact, S3 delivers performance comparable to or better than AdamW with \textbf{$2 \times$} the training steps, establishing its efficacy in both efficiency and final task performance.

Qi Wang, Hanyang Peng, Yue Yu

Mixture-of-Experts (MoE) models enable scalable performance by activating large parameter sets sparsely, minimizing computational overhead. To mitigate the prohibitive cost of training MoEs from scratch, recent work employs upcycling, reusing a single pre-trained dense model by replicating its feed-forward network (FFN) layers into experts. However, this limits expert diversity, as all experts originate from a single pre-trained dense model. This paper addresses this limitation by constructing powerful MoE models using experts sourced from multiple identically-architected but disparate pre-trained models (e.g., Qwen2.5-Coder and Qwen2). A key challenge lies in the fact that these source models occupy disparate, dissonant regions of the parameter space, making direct upcycling prone to severe performance degradation. To overcome this, we propose Symphony-MoE, a novel two-stage framework designed to harmonize these models into a single, coherent expert mixture. First, we establish this harmony in a training-free manner: we construct a shared backbone via a layer-aware fusion strategy and, crucially, alleviate parameter misalignment among experts using activation-based functional alignment. Subsequently, a stage of post-training coordinates the entire architecture. Experiments demonstrate that our method successfully integrates experts from heterogeneous sources, achieving an MoE model that significantly surpasses baselines in multi-domain tasks and out-of-distribution generalization.

Hanyang Peng, Shuang Qin, Yue Yu et al.

Adam is widely recognized as one of the most effective optimizers for training deep neural networks (DNNs). Despite its remarkable empirical success, its theoretical convergence analysis remains unsatisfactory. Existing works predominantly interpret Adam as a preconditioned stochastic gradient descent with momentum (SGDM), formulated as $\bm{x}_{t+1} = \bm{x}_t - \frac{γ_t}{{\sqrt{\bm{v}_t}+ε}} \circ \bm{m}_t$. This perspective necessitates strong assumptions and intricate techniques, resulting in lengthy and opaque convergence proofs that are difficult to verify and extend. In contrast, we propose a novel interpretation by treating Adam as a sign-like optimizer, expressed as $\bm{x}_{t+1} = \bm{x}_t - γ_t \frac{|\bm{m}_t|}{{\sqrt{\bm{v}_t}+ε}} \circ {\rm Sign}(\bm{m}_t)$. This reformulation significantly simplifies the convergence analysis. For the first time, with some mild conditions, we prove that Adam achieves the optimal rate of ${\cal O}(\frac{1}{T^{\sfrac{1}{4}}})$ rather than the previous ${\cal O} \left(\frac{\ln T}{T^{\sfrac{1}{4}}}\right)$ under weak assumptions of the generalized $p$-affine variance and $(L_0, L_1, q)$-smoothness, without dependence on the model dimensionality or the numerical stability parameter $ε$. Additionally, our theoretical analysis provides new insights into the role of momentum as a key factor ensuring convergence and offers practical guidelines for tuning learning rates in Adam, further bridging the gap between theory and practice.

Hanyang Peng, Shiqi Yu

Four-variable-independent-regression localization losses, such as Smooth-$\ell_1$ Loss, are used by default in modern detectors. Nevertheless, this kind of loss is oversimplified so that it is inconsistent with the final evaluation metric, intersection over union (IoU). Directly employing the standard IoU is also not infeasible, since the constant-zero plateau in the case of non-overlapping boxes and the non-zero gradient at the minimum may make it not trainable. Accordingly, we propose a systematic method to address these problems. Firstly, we propose a new metric, the extended IoU (EIoU), which is well-defined when two boxes are not overlapping and reduced to the standard IoU when overlapping. Secondly, we present the convexification technique (CT) to construct a loss on the basis of EIoU, which can guarantee the gradient at the minimum to be zero. Thirdly, we propose a steady optimization technique (SOT) to make the fractional EIoU loss approaching the minimum more steadily and smoothly. Fourthly, to fully exploit the capability of the EIoU based loss, we introduce an interrelated IoU-predicting head to further boost localization accuracy. With the proposed contributions, the new method incorporated into Faster R-CNN with ResNet50+FPN as the backbone yields \textbf{4.2 mAP} gain on VOC2007 and \textbf{2.3 mAP} gain on COCO2017 over the baseline Smooth-$\ell_1$ Loss, at almost \textbf{no training and inferencing computational cost}. Specifically, the stricter the metric is, the more notable the gain is, improving \textbf{8.2 mAP} on VOC2007 and \textbf{5.4 mAP} on COCO2017 at metric $AP_{90}$.

Yuantao Feng, Shiqi Yu, Hanyang Peng et al.

Face detection is to search all the possible regions for faces in images and locate the faces if there are any. Many applications including face recognition, facial expression recognition, face tracking and head-pose estimation assume that both the location and the size of faces are known in the image. In recent decades, researchers have created many typical and efficient face detectors from the Viola-Jones face detector to current CNN-based ones. However, with the tremendous increase in images and videos with variations in face scale, appearance, expression, occlusion and pose, traditional face detectors are challenged to detect various "in the wild" faces. The emergence of deep learning techniques brought remarkable breakthroughs to face detection along with the price of a considerable increase in computation. This paper introduces representative deep learning-based methods and presents a deep and thorough analysis in terms of accuracy and efficiency. We further compare and discuss the popular and challenging datasets and their evaluation metrics. A comprehensive comparison of several successful deep learning-based face detectors is conducted to uncover their efficiency using two metrics: FLOPs and latency. The paper can guide to choose appropriate face detectors for different applications and also to develop more efficient and accurate detectors.