Rong Xiao

Yu Wan, Baosong Yang, Dayiheng Liu et al.

Attention mechanism has become the dominant module in natural language processing models. It is computationally intensive and depends on massive power-hungry multiplications. In this paper, we rethink variants of attention mechanism from the energy consumption aspects. After reaching the conclusion that the energy costs of several energy-friendly operations are far less than their multiplication counterparts, we build a novel attention model by replacing multiplications with either selective operations or additions. Empirical results on three machine translation tasks demonstrate that the proposed model, against the vanilla one, achieves competitable accuracy while saving 99\% and 66\% energy during alignment calculation and the whole attention procedure. Code is available at: https://github.com/NLP2CT/E-Att.

Yu Liang, Yufeng Zhang, Shiliang Zhang et al.

Backward-compatible training circumvents the need for expensive updates to the old gallery database when deploying an advanced new model in the retrieval system. Previous methods achieved backward compatibility by aligning prototypes of the new model with the old one, yet they often overlooked the distribution of old features, limiting their effectiveness when the low quality of the old model results in a weakly feature discriminability. Instance-based methods like L2 regression take into account the distribution of old features but impose strong constraints on the performance of the new model itself. In this paper, we propose MixBCT, a simple yet highly effective backward-compatible training method that serves as a unified framework for old models of varying qualities. We construct a single loss function applied to mixed old and new features to facilitate backward-compatible training, which adaptively adjusts the constraint domain for new features based on the distribution of old features. We conducted extensive experiments on the large-scale face recognition datasets MS1Mv3 and IJB-C to verify the effectiveness of our method. The experimental results clearly demonstrate its superiority over previous methods. Code is available at https://github.com/yuleung/MixBCT .

Yun Yi, Haokui Zhang, Rong Xiao et al.

As more deep learning models are being applied in real-world applications, there is a growing need for modeling and learning the representations of neural networks themselves. An efficient representation can be used to predict target attributes of networks without the need for actual training and deployment procedures, facilitating efficient network deployment and design. Recently, inspired by the success of Transformer, some Transformer-based representation learning frameworks have been proposed and achieved promising performance in handling cell-structured models. However, graph neural network (GNN) based approaches still dominate the field of learning representation for the entire network. In this paper, we revisit Transformer and compare it with GNN to analyse their different architecture characteristics. We then propose a modified Transformer-based universal neural network representation learning model NAR-Former V2. It can learn efficient representations from both cell-structured networks and entire networks. Specifically, we first take the network as a graph and design a straightforward tokenizer to encode the network into a sequence. Then, we incorporate the inductive representation learning capability of GNN into Transformer, enabling Transformer to generalize better when encountering unseen architecture. Additionally, we introduce a series of simple yet effective modifications to enhance the ability of the Transformer in learning representation from graph structures. Our proposed method surpasses the GNN-based method NNLP by a significant margin in latency estimation on the NNLQP dataset. Furthermore, regarding accuracy prediction on the NASBench101 and NASBench201 datasets, our method achieves highly comparable performance to other state-of-the-art methods.

Youze Huang, Penghui Ruan, Bojia Zi et al.

Recent advances in diffusion models have brought remarkable progress in image and video editing, yet some tasks remain underexplored. In this paper, we introduce a new task, Object Retexture, which transfers local textures from a reference object to a target object in images or videos. To perform this task, a straightforward solution is to use ControlNet conditioned on the source structure and the reference texture. However, this approach suffers from limited controllability for two reasons: conditioning on the raw reference image introduces unwanted structural information, and it fails to disentangle the visual texture and structure information of the source. To address this problem, we propose Refaçade, a method that consists of two key designs to achieve precise and controllable texture transfer in both images and videos. First, we employ a texture remover trained on paired textured/untextured 3D mesh renderings to remove appearance information while preserving the geometry and motion of source videos. Second, we disrupt the reference global layout using a jigsaw permutation, encouraging the model to focus on local texture statistics rather than the global layout of the object. Extensive experiments demonstrate superior visual quality, precise editing, and controllability, outperforming strong baselines in both quantitative and human evaluations. Code is available at https://github.com/fishZe233/Refacade.

Haokui Zhang, Congyang Ou, Dawei Yan et al.

Recently, reducing redundant visual tokens in vision-language models (VLMs) to accelerate VLM inference has emerged as a hot topic. However, most existing methods rely on heuristics constructed based on inter-visual-token similarity or cross-modal visual-text similarity, which gives rise to certain limitations in compression performance and practical deployment. In contrast, we propose PIO-FVLM from the perspective of inference objectives, which transforms visual token compression into preserving output result invariance and selects tokens primarily by their importance to this goal. Specially, vision tokens are reordered with the guidance of token-level gradient saliency generated by our designed layer-local proxy loss, a coarse constraint from the current layer to the final result. Then the most valuable vision tokens are selected following the non-maximum suppression (NMS) principle. The proposed PIO-FVLM is training-free and compatible with FlashAttention, friendly to practical application and deployment. It can be deployed independently as an encoder-free method, or combined with encoder compression approaches like VisionZip for use as an encoder-involved method. On LLaVA-Next-7B, PIO-FVLM retains just 11.1% of visual tokens but maintains 97.2% of the original performance, with a 2.67$\times$ prefill speedup, 2.11$\times$ inference speedup, 6.22$\times$ lower FLOPs, and 6.05$\times$ reduced KV Cache overhead. Our code is available at https://github.com/ocy1/PIO-FVLM.

Xianbiao Qi, Marco Chen, Jiaquan Ye et al.

The Muon optimizer has recently attracted considerable attention for its strong empirical performance and use of orthogonalized updates on matrix-shaped parameters, yet its underlying mechanisms and relationship to adaptive optimizers such as Adam remain insufficiently understood. In this work, we aim to address these questions through a unified spectral perspective. Specifically, we view Muon as the p = 0 endpoint of a family of spectral transformations of the form U \boldsymbolΣ^{p} V' , and consider additional variants with p = 1/2 , p = 1/4 , and p = 1 . These transformations are applied to both first-moment updates, as in momentum SGD, and to root-mean-square (RMS) normalized gradient updates as in Adam. To enable efficient computation, we develop a coupled Newton iteration that avoids explicit singular value decomposition. Across controlled experiments, we find that RMS-normalized updates yield more stable optimization than first-moment updates. Moreover, while spectral compression provides strong stabilization benefits under first-moment updates, the Muon update (p = 0) does not consistently outperform Adam. These results suggest that Muon is best understood as an effective form of spectral normalization, but not a universally superior optimization method. Our source code will be released at https://github.com/Ocram7/BeyondMuon.

Guoting Wei, Xia Yuan, Yu Liu et al.

Aerial object detection plays a crucial role in numerous applications. However, most existing methods focus on detecting predefined object categories, limiting their applicability in real-world open scenarios. In this paper, we extend aerial object detection to open scenarios through image-text collaboration and propose RT-OVAD, the first real-time open-vocabulary detector for aerial scenes. Specifically, we first introduce an image-to-text alignment loss to replace the conventional category regression loss, thereby eliminating category constraints. Next, we propose a lightweight image-text collaboration strategy comprising an image-text collaboration encoder and a text-guided decoder. The encoder simultaneously enhances visual features and refines textual embeddings, while the decoder guides object queries to focus on class-relevant image features. This design further improves detection accuracy without incurring significant computational overhead. Extensive experiments demonstrate that RT-OVAD consistently outperforms existing state-of-the-art methods across open-vocabulary, zero-shot, and traditional closed-set detection tasks. For instance, on the open-vocabulary aerial detection benchmarks DIOR, DOTA-v2.0, and LAE-80C, RT-OVAD achieves 87.7 AP$_{50}$, 53.8 mAP, and 23.7 mAP, respectively, surpassing the previous state-of-the-art (LAE-DINO) by 2.2, 7.0, and 3.5 points. In addition, RT-OVAD achieves an inference speed of 34 FPS on an RTX 4090 GPU, approximately three times faster than LAE-DINO (10 FPS), meeting the real-time detection requirements of diverse applications. The code will be released at https://github.com/GT-Wei/RT-OVAD.

Jiong Cai, Yong Jiang, Yue Zhang et al.

Discovering the intended items of user queries from a massive repository of items is one of the main goals of an e-commerce search system. Relevance prediction is essential to the search system since it helps improve performance. When online serving a relevance model, the model is required to perform fast and accurate inference. Currently, the widely used models such as Bi-encoder and Cross-encoder have their limitations in accuracy or inference speed respectively. In this work, we propose a novel model called the Entity-Based Relevance Model (EBRM). We identify the entities contained in an item and decompose the QI (query-item) relevance problem into multiple QE (query-entity) relevance problems; we then aggregate their results to form the QI prediction using a soft logic formulation. The decomposition allows us to use a Cross-encoder QE relevance module for high accuracy as well as cache QE predictions for fast online inference. Utilizing soft logic makes the prediction procedure interpretable and intervenable. We also show that pretraining the QE module with auto-generated QE data from user logs can further improve the overall performance. The proposed method is evaluated on labeled data from e-commerce websites. Empirical results show that it achieves promising improvements with computation efficiency.

Chao Tao, Aoran Hu, Rong Xiao et al.

Data-driven deep learning methods have shown great potential in cropland mapping. However, due to multiple factors such as attributes of cropland (topography, climate, crop type) and imaging conditions (viewing angle, illumination, scale), croplands under different scenes demonstrate a great domain gap. This makes it difficult for models trained in the specific scenes to directly generalize to other scenes. A common way to handle this problem is through the "Pretrain+Fine-tuning" paradigm. Unfortunately, considering the variety of features of cropland that are affected by multiple factors, it is hardly to handle the complex domain gap between pre-trained data and target data using only sparse fine-tuned samples as general constraints. Moreover, as the number of model parameters grows, fine-tuning is no longer an easy and low-cost task. With the emergence of prompt learning via visual foundation models, the "Pretrain+Prompting" paradigm redesigns the optimization target by introducing individual prompts for each single sample. This simplifies the domain adaption from generic to specific scenes during model reasoning processes. Therefore, we introduce the "Pretrain+Prompting" paradigm to interpreting cropland scenes and design the auto-prompting (APT) method based on freely available global land cover product. It can achieve a fine-grained adaptation process from generic scenes to specialized cropland scenes without introducing additional label costs. To our best knowledge, this work pioneers the exploration of the domain adaption problems for cropland mapping under prompt learning perspectives. Our experiments using two sub-meter cropland datasets from southern and northern China demonstrated that the proposed method via visual foundation models outperforms traditional supervised learning and fine-tuning approaches in the field of remote sensing.

Xuantong Liu, Shaozhe Hao, Xianbiao Qi et al.

The success of autoregressive (AR) language models in text generation has inspired the computer vision community to adopt Large Language Models (LLMs) for image generation. However, considering the essential differences between text and image modalities, the design space of language models for image generation remains underexplored. We observe that image tokens exhibit greater randomness compared to text tokens, which presents challenges when training with token prediction. Nevertheless, AR models demonstrate their potential by effectively learning patterns even from a seemingly suboptimal optimization problem. Our analysis also reveals that while all models successfully grasp the importance of local information in image generation, smaller models struggle to capture the global context. In contrast, larger models showcase improved capabilities in this area, helping to explain the performance gains achieved when scaling up model size. We further elucidate the design space of language models for vision generation, including tokenizer choice, model choice, model scalability, vocabulary design, and sampling strategy through extensive comparative experiments. Our work is the first to analyze the optimization behavior of language models in vision generation, and we believe it can inspire more effective designs when applying LMs to other domains. Finally, our elucidated language model for image generation, termed as ELM, achieves state-of-the-art performance on the ImageNet 256*256 benchmark. The code is available at https://github.com/Pepperlll/LMforImageGeneration.git.

Wei He, Shangzhi Zhang, Chun-Guang Li et al.

Spectral clustering, as a popular tool for data clustering, requires an eigen-decomposition step on a given affinity to obtain the spectral embedding. Nevertheless, such a step suffers from the lack of generalizability and scalability. Moreover, the obtained spectral embeddings can hardly provide a good approximation to the ground-truth partition and thus a k-means step is adopted to quantize the embedding. In this paper, we propose a simple yet effective scalable and generalizable approach, called Neural Normalized Cut (NeuNcut), to learn the clustering membership for spectral clustering directly. In NeuNcut, we properly reparameterize the unknown cluster membership via a neural network, and train the neural network via stochastic gradient descent with a properly relaxed normalized cut loss. As a result, our NeuNcut enjoys a desired generalization ability to directly infer clustering membership for out-of-sample unseen data and hence brings us an efficient way to handle clustering task with ultra large-scale data. We conduct extensive experiments on both synthetic data and benchmark datasets and experimental results validate the effectiveness and the superiority of our approach. Our code is available at: https://github.com/hewei98/NeuNcut.

Xianghan Meng, Zhiyuan Huang, Wei He et al.

Subspace clustering is a classical unsupervised learning task, built on a basic assumption that high-dimensional data can be approximated by a union of subspaces (UoS). Nevertheless, the real-world data are often deviating from the UoS assumption. To address this challenge, state-of-the-art deep subspace clustering algorithms attempt to jointly learn UoS representations and self-expressive coefficients. However, the general framework of the existing algorithms suffers from a catastrophic feature collapse and lacks a theoretical guarantee to learn desired UoS representation. In this paper, we present a Principled fRamewOrk for Deep Subspace Clustering (PRO-DSC), which is designed to learn structured representations and self-expressive coefficients in a unified manner. Specifically, in PRO-DSC, we incorporate an effective regularization on the learned representations into the self-expressive model, prove that the regularized self-expressive model is able to prevent feature space collapse, and demonstrate that the learned optimal representations under certain condition lie on a union of orthogonal subspaces. Moreover, we provide a scalable and efficient approach to implement our PRO-DSC and conduct extensive experiments to verify our theoretical findings and demonstrate the superior performance of our proposed deep subspace clustering approach. The code is available at https://github.com/mengxianghan123/PRO-DSC.

Wei He, Xianghan Meng, Zhiyuan Huang et al.

Generalized Category Discovery (GCD) aims to identify both known and unknown categories, with only partial labels given for the known categories, posing a challenging open-set recognition problem. State-of-the-art approaches for GCD task are usually built on multi-modality representation learning, which is heavily dependent upon inter-modality alignment. However, few of them cast a proper intra-modality alignment to generate a desired underlying structure of representation distributions. In this paper, we propose a novel and effective multi-modal representation learning framework for GCD via Semi-Supervised Rate Reduction, called SSR$^2$-GCD, to learn cross-modality representations with desired structural properties based on emphasizing to properly align intra-modality relationships. Moreover, to boost knowledge transfer, we integrate prompt candidates by leveraging the inter-modal alignment offered by Vision Language Models. We conduct extensive experiments on generic and fine-grained benchmark datasets demonstrating superior performance of our approach.

Guoting Wei, Yu Liu, Xia Yuan et al.

In recent years, language-guided open-set aerial object detection has gained significant attention due to its better alignment with real-world application needs. However, due to limited datasets, most existing language-guided methods primarily focus on vocabulary-level descriptions, which fail to meet the demands of fine-grained open-world detection. To address this limitation, we propose constructing a large-scale language-guided open-set aerial detection dataset, encompassing three levels of language guidance: from words to phrases, and ultimately to sentences. Centered around an open-source large vision-language model and integrating image-operation-based preprocessing with BERT-based postprocessing, we present the OS-W2S Label Engine, an automatic annotation pipeline capable of handling diverse scene annotations for aerial images. Using this label engine, we expand existing aerial detection datasets with rich textual annotations and construct a novel benchmark dataset, called MI-OAD, addressing the limitations of current remote sensing grounding data and enabling effective language-guided open-set aerial detection. Specifically, MI-OAD contains 163,023 images and 2 million image-caption pairs, approximately 40 times larger than comparable datasets. To demonstrate the effectiveness and quality of MI-OAD, we evaluate three representative tasks. On language-guided open-set aerial detection, training on MI-OAD lifts Grounding DINO by +31.1 AP$_{50}$ and +34.7 Recall@10 with sentence-level inputs under zero-shot transfer. Moreover, using MI-OAD for pre-training yields state-of-the-art performance on multiple existing open-vocabulary aerial detection and remote sensing visual grounding benchmarks, validating both the effectiveness of the dataset and the high quality of its OS-W2S annotations. More details are available at https://github.com/GT-Wei/MI-OAD.

Yuanjian Wang, Yufei Deng, Rong Xiao et al.

Event cameras are neuromorphic vision sensors that asynchronously capture changes in logarithmic brightness changes, offering significant advantages such as low latency, low power consumption, low bandwidth, and high dynamic range. While these characteristics make them ideal for high-speed scenarios, reconstructing geometrically consistent and photometrically accurate 3D representations from event data remains fundamentally challenging. Current event-based Neural Radiance Fields (NeRF) methods partially address these challenges but suffer from persistent artifacts caused by aggressive network learning in early stages and the inherent noise of event cameras. To overcome these limitations, we present SaENeRF, a novel self-supervised framework that effectively suppresses artifacts and enables 3D-consistent, dense, and photorealistic NeRF reconstruction of static scenes solely from event streams. Our approach normalizes predicted radiance variations based on accumulated event polarities, facilitating progressive and rapid learning for scene representation construction. Additionally, we introduce regularization losses specifically designed to suppress artifacts in regions where photometric changes fall below the event threshold and simultaneously enhance the light intensity difference of non-zero events, thereby improving the visual fidelity of the reconstructed scene. Extensive qualitative and quantitative experiments demonstrate that our method significantly reduces artifacts and achieves superior reconstruction quality compared to existing methods. The code is available at https://github.com/Mr-firework/SaENeRF.

Hanling Yi, Feng Lin, Mao Luo et al.

Recent advances in Multimodal Large Language Models (MLLMs) have enabled open-ended object recognition, yet they struggle with fine-grained tasks. In contrast, CLIP-style models excel at fine-grained recognition but lack broad coverage of general object categories. To bridge this gap, we propose \textbf{HyMOR}, a \textbf{Hy}brid \textbf{M}ulti-granularity open-ended \textbf{O}bject \textbf{R}ecognition framework that integrates an MLLM with a CLIP model. In HyMOR, the MLLM performs open-ended and coarse-grained object recognition, while the CLIP model specializes in fine-grained identification of domain-specific objects such as animals and plants. This hybrid design enables accurate object understanding across multiple semantic granularities, serving as a robust perceptual foundation for downstream multi-modal content generation and interactive gameplay. To support evaluation in content-rich and educational scenarios, we introduce TBO (TextBook Objects), a dataset containing 20,942 images annotated with 8,816 object categories extracted from textbooks. Extensive experiments demonstrate that HyMOR narrows the fine-grained recognition gap with CLIP to 0.2\% while improving general object recognition by 2.5\% over a baseline MLLM, measured by average Sentence-BERT (SBert) similarity. Overall, HyMOR achieves a 23.2\% improvement in average SBert across all evaluated datasets, highlighting its effectiveness in enabling accurate perception for multi-modal game content generation and interactive learning applications.

Feiyu An, Yufei Deng, Zihui Zhang et al.

Achieving sharp 3D reconstruction from motion-blurred images alone becomes challenging, motivating recent methods to incorporate event cameras, benefiting from microsecond temporal resolution. However, they suffer from residual artifacts and blurry texture details due to misleading supervision from inaccurate event double integral priors and noisy, blurry events. In this study, we propose EvFlow-GS, a unified framework that leverages event streams and optical flow to optimize an end-to-end learnable double integral (LDI), camera poses, and 3D Gaussian Splatting (3DGS) jointly on-the-fly. Specifically, we first extract edge information from the events using optical flow and then formulate a novel event-based loss applied separately to different modules. Additionally, we exploit a novel event-residual prior to strengthen the supervision of intensity changes between images rendered from 3DGS. Finally, we integrate the outputs of both 3DGS and LDI into a joint loss, enabling their optimization to mutually facilitate each other. Experiments demonstrate the leading performance of our EvFlow-GS.

Marco Chen, Xianbiao Qi, Yelin He et al.

In this work, we revisit Transformer optimization through the lens of second-order geometry and establish a direct connection between architectural design, activation scale, the Hessian matrix, and the maximum tolerable learning rate. We introduce a simple normalization strategy, termed SimpleNorm, which stabilizes intermediate activation scales by construction. Then, by analyzing the Hessian of the loss with respect to network activations, we theoretically show that SimpleNorm significantly reduces the spectral norm of the Hessian, thereby permitting larger stable learning rates. We validate our theoretical findings through extensive experiments on large GPT models at parameter scales 1B, 1.4B, 7B and 8B. Empirically, SimpleGPT, our SimpleNorm-based network, tolerates learning rates 3$\times$-10$\times$ larger than standard convention, consistently demonstrates strong optimization stability, and achieves substantially better performance than well-established baselines. Specifically, when training 7B-scale models for 60K steps, SimpleGPT achieves a training loss that is 0.08 lower than that of LLaMA2 with QKNorm, reducing the loss from 2.290 to 2.208. Our source code will be released at https://github.com/Ocram7/SimpleGPT.

Yuze Wang, Rong Xiao, Haifeng Li et al.

In remote sensing scene classification, leveraging the transfer methods with well-trained optical models is an efficient way to overcome label scarcity. However, cloud contamination leads to optical information loss and significant impacts on feature distribution, challenging the reliability and stability of transferred target models. Common solutions include cloud removal for optical data or directly using Synthetic aperture radar (SAR) data in the target domain. However, cloud removal requires substantial auxiliary data for support and pre-training, while directly using SAR disregards the unobstructed portions of optical data. This study presents a scene classification transfer method that synergistically combines multi-modality data, which aims to transfer the source domain model trained on cloudfree optical data to the target domain that includes both cloudy optical and SAR data at low cost. Specifically, the framework incorporates two parts: (1) the collaborative transfer strategy, based on knowledge distillation, enables the efficient prior knowledge transfer across heterogeneous data; (2) the information regulation mechanism (IRM) is proposed to address the modality imbalance issue during transfer. It employs auxiliary models to measure the contribution discrepancy of each modality, and automatically balances the information utilization of modalities during the target model learning process at the sample-level. The transfer experiments were conducted on simulated and real cloud datasets, demonstrating the superior performance of the proposed method compared to other solutions in cloud-covered scenarios. We also verified the importance and limitations of IRM, and further discussed and visualized the modality imbalance problem during the model transfer. Codes are available at https://github.com/wangyuze-csu/ESCCS

Yihao Chen, Xin Tang, Xianbiao Qi et al.

Graph Neural Networks (GNNs) have attracted considerable attention and have emerged as a new promising paradigm to process graph-structured data. GNNs are usually stacked to multiple layers and the node representations in each layer are computed through propagating and aggregating the neighboring node features with respect to the graph. By stacking to multiple layers, GNNs are able to capture the long-range dependencies among the data on the graph and thus bring performance improvements. To train a GNN with multiple layers effectively, some normalization techniques (e.g., node-wise normalization, batch-wise normalization) are necessary. However, the normalization techniques for GNNs are highly task-relevant and different application tasks prefer to different normalization techniques, which is hard to know in advance. To tackle this deficiency, in this paper, we propose to learn graph normalization by optimizing a weighted combination of normalization techniques at four different levels, including node-wise normalization, adjacency-wise normalization, graph-wise normalization, and batch-wise normalization, in which the adjacency-wise normalization and the graph-wise normalization are newly proposed in this paper to take into account the local structure and the global structure on the graph, respectively. By learning the optimal weights, we are able to automatically select a single best or a best combination of multiple normalizations for a specific task. We conduct extensive experiments on benchmark datasets for different tasks, including node classification, link prediction, graph classification and graph regression, and confirm that the learned graph normalization leads to competitive results and that the learned weights suggest the appropriate normalization techniques for the specific task. Source code is released here https://github.com/cyh1112/GraphNormalization.

Zhihong Chen, Rong Xiao, Chenliang Li et al.

Most of ranking models are trained only with displayed items (most are hot items), but they are utilized to retrieve items in the entire space which consists of both displayed and non-displayed items (most are long-tail items). Due to the sample selection bias, the long-tail items lack sufficient records to learn good feature representations, i.e. data sparsity and cold start problems. The resultant distribution discrepancy between displayed and non-displayed items would cause poor long-tail performance. To this end, we propose an entire space adaptation model (ESAM) to address this problem from the perspective of domain adaptation (DA). ESAM regards displayed and non-displayed items as source and target domains respectively. Specifically, we design the attribute correlation alignment that considers the correlation between high-level attributes of the item to achieve distribution alignment. Furthermore, we introduce two effective regularization strategies, i.e. \textit{center-wise clustering} and \textit{self-training} to improve DA process. Without requiring any auxiliary information and auxiliary domains, ESAM transfers the knowledge from displayed items to non-displayed items for alleviating the distribution inconsistency. Experiments on two public datasets and a large-scale industrial dataset collected from Taobao demonstrate that ESAM achieves state-of-the-art performance, especially in the long-tail space. Besides, we deploy ESAM to the Taobao search engine, leading to significant improvement on online performance. The code is available at \url{https://github.com/A-bone1/ESAM.git}

Wenwen Yu, Ning Lu, Xianbiao Qi et al.

Computer vision with state-of-the-art deep learning models has achieved huge success in the field of Optical Character Recognition (OCR) including text detection and recognition tasks recently. However, Key Information Extraction (KIE) from documents as the downstream task of OCR, having a large number of use scenarios in real-world, remains a challenge because documents not only have textual features extracting from OCR systems but also have semantic visual features that are not fully exploited and play a critical role in KIE. Too little work has been devoted to efficiently make full use of both textual and visual features of the documents. In this paper, we introduce PICK, a framework that is effective and robust in handling complex documents layout for KIE by combining graph learning with graph convolution operation, yielding a richer semantic representation containing the textual and visual features and global layout without ambiguity. Extensive experiments on real-world datasets have been conducted to show that our method outperforms baselines methods by significant margins. Our code is available at https://github.com/wenwenyu/PICK-pytorch.

Ning Lu, Wenwen Yu, Xianbiao Qi et al.

Attention-based scene text recognizers have gained huge success, which leverages a more compact intermediate representation to learn 1d- or 2d- attention by a RNN-based encoder-decoder architecture. However, such methods suffer from attention-drift problem because high similarity among encoded features leads to attention confusion under the RNN-based local attention mechanism. Moreover, RNN-based methods have low efficiency due to poor parallelization. To overcome these problems, we propose the MASTER, a self-attention based scene text recognizer that (1) not only encodes the input-output attention but also learns self-attention which encodes feature-feature and target-target relationships inside the encoder and decoder and (2) learns a more powerful and robust intermediate representation to spatial distortion, and (3) owns a great training efficiency because of high training parallelization and a high-speed inference because of an efficient memory-cache mechanism. Extensive experiments on various benchmarks demonstrate the superior performance of our MASTER on both regular and irregular scene text. Pytorch code can be found at https://github.com/wenwenyu/MASTER-pytorch, and Tensorflow code can be found at https://github.com/jiangxiluning/MASTER-TF.

Yinfu Feng, Yanjing Wu, Rong Xiao et al.

Item-to-Item (I2I) recommendation models are widely used in real-world systems due to their scalability, real-time capabilities, and high recommendation quality. Research to enhance I2I performance focuses on two directions: 1) model-centric approaches, which adopt deeper architectures but risk increased computational costs and deployment complexity, and 2) data-centric methods, which refine training data without altering models, offering cost-effectiveness but struggling with data sparsity and noise. To address these challenges, we propose LLM-I2I, a data-centric framework leveraging Large Language Models (LLMs) to mitigate data quality issues. LLM-I2I includes (1) an LLM-based generator that synthesizes user-item interactions for long-tail items, alleviating data sparsity, and (2) an LLM-based discriminator that filters noisy interactions from real and synthetic data. The refined data is then fused to train I2I models. Evaluated on industry (AEDS) and academic (ARD) datasets, LLM-I2I consistently improves recommendation accuracy, particularly for long-tail items. Deployed on a large-scale cross-border e-commerce platform, it boosts recall number (RN) by 6.02% and gross merchandise value (GMV) by 1.22% over existing I2I models. This work highlights the potential of LLMs in enhancing data-centric recommendation systems without modifying model architectures.

Hanling Yi, Feng Lin, Hongbin Li et al.

This research aims to accelerate the inference speed of large language models (LLMs) with billions of parameters. We propose \textbf{S}mart \textbf{P}arallel \textbf{A}uto-\textbf{C}orrect d\textbf{E}coding (SPACE), an innovative approach designed for achieving lossless acceleration of LLMs. By integrating semi-autoregressive inference and speculative decoding capabilities, SPACE uniquely enables autoregressive LLMs to parallelize token generation and verification. This is realized through a specialized semi-autoregressive supervised fine-tuning process that equips existing LLMs with the ability to simultaneously predict multiple tokens. Additionally, an auto-correct decoding algorithm facilitates the simultaneous generation and verification of token sequences within a single model invocation. Through extensive experiments on a range of LLMs, SPACE has demonstrated inference speedup ranging from 2.7x-4.0x on HumanEval-X while maintaining output quality.

Bojia Zi, Penghui Ruan, Marco Chen et al.

Recent advancements in video generation have spurred the development of video editing techniques, which can be divided into inversion-based and end-to-end methods. However, current video editing methods still suffer from several challenges. Inversion-based methods, though training-free and flexible, are time-consuming during inference, struggle with fine-grained editing instructions, and produce artifacts and jitter. On the other hand, end-to-end methods, which rely on edited video pairs for training, offer faster inference speeds but often produce poor editing results due to a lack of high-quality training video pairs. In this paper, to close the gap in end-to-end methods, we introduce Señorita-2M, a high-quality video editing dataset. Señorita-2M consists of approximately 2 millions of video editing pairs. It is built by crafting four high-quality, specialized video editing models, each crafted and trained by our team to achieve state-of-the-art editing results. We also propose a filtering pipeline to eliminate poorly edited video pairs. Furthermore, we explore common video editing architectures to identify the most effective structure based on current pre-trained generative model. Extensive experiments show that our dataset can help to yield remarkably high-quality video editing results. More details are available at https://senorita-2m-dataset.github.io.

Feng Lin, Hanling Yi, Hongbin Li et al.

Large language models (LLMs) commonly employ autoregressive generation during inference, leading to high memory bandwidth demand and consequently extended latency. To mitigate this inefficiency, we present Bi-directional Tuning for lossless Acceleration (BiTA), an innovative method expediting LLMs via streamlined semi-autoregressive generation and draft verification. Inspired by the concept of prompt tuning, we enhance LLMs with a parameter-efficient design called bi-directional tuning for the capability in semi-autoregressive generation. Employing efficient tree-based decoding, the models perform draft candidate generation and verification in parallel, ensuring outputs identical to their autoregressive counterparts under greedy sampling. BiTA serves as a lightweight plug-in module, seamlessly boosting the inference efficiency of existing LLMs without requiring additional assistance models or incurring significant extra memory costs. Applying the proposed BiTA, LLaMA-2-70B-Chat achieves a 2.7$\times$ speedup on the MT-Bench benchmark. Extensive experiments confirm our method surpasses state-of-the-art acceleration techniques.

Bojia Zi, Weixuan Peng, Xianbiao Qi et al.

Recent advances in video diffusion models have driven rapid progress in video editing techniques. However, video object removal, a critical subtask of video editing, remains challenging due to issues such as hallucinated objects and visual artifacts. Furthermore, existing methods often rely on computationally expensive sampling procedures and classifier-free guidance (CFG), resulting in slow inference. To address these limitations, we propose MiniMax-Remover, a novel two-stage video object removal approach. Motivated by the observation that text condition is not best suited for this task, we simplify the pretrained video generation model by removing textual input and cross-attention layers, resulting in a more lightweight and efficient model architecture in the first stage. In the second stage, we distilled our remover on successful videos produced by the stage-1 model and curated by human annotators, using a minimax optimization strategy to further improve editing quality and inference speed. Specifically, the inner maximization identifies adversarial input noise ("bad noise") that makes failure removals, while the outer minimization step trains the model to generate high-quality removal results even under such challenging conditions. As a result, our method achieves a state-of-the-art video object removal results with as few as 6 sampling steps and doesn't rely on CFG, significantly improving inference efficiency. Extensive experiments demonstrate the effectiveness and superiority of MiniMax-Remover compared to existing methods. Codes and Videos are available at: https://minimax-remover.github.io.

Yufei Deng, Yuanjian Wang, Rong Xiao et al.

While 3D Gaussian Splatting (3D-GS) achieves photorealistic novel view synthesis, its performance degrades with motion blur. In scenarios with rapid motion or low-light conditions, existing RGB-based deblurring methods struggle to model camera pose and radiance changes during exposure, reducing reconstruction accuracy. Event cameras, capturing continuous brightness changes during exposure, can effectively assist in modeling motion blur and improving reconstruction quality. Therefore, we propose Event-driven Bundle Adjusted Deblur Gaussian Splatting (EBAD-Gaussian), which reconstructs sharp 3D Gaussians from event streams and severely blurred images. This method jointly learns the parameters of these Gaussians while recovering camera motion trajectories during exposure time. Specifically, we first construct a blur loss function by synthesizing multiple latent sharp images during the exposure time, minimizing the difference between real and synthesized blurred images. Then we use event stream to supervise the light intensity changes between latent sharp images at any time within the exposure period, supplementing the light intensity dynamic changes lost in RGB images. Furthermore, we optimize the latent sharp images at intermediate exposure times based on the event-based double integral (EDI) prior, applying consistency constraints to enhance the details and texture information of the reconstructed images. Extensive experiments on synthetic and real-world datasets show that EBAD-Gaussian can achieve high-quality 3D scene reconstruction under the condition of blurred images and event stream inputs.

Shaozhe Hao, Xuantong Liu, Xianbiao Qi et al.

We introduce BiGR, a novel conditional image generation model using compact binary latent codes for generative training, focusing on enhancing both generation and representation capabilities. BiGR is the first conditional generative model that unifies generation and discrimination within the same framework. BiGR features a binary tokenizer, a masked modeling mechanism, and a binary transcoder for binary code prediction. Additionally, we introduce a novel entropy-ordered sampling method to enable efficient image generation. Extensive experiments validate BiGR's superior performance in generation quality, as measured by FID-50k, and representation capabilities, as evidenced by linear-probe accuracy. Moreover, BiGR showcases zero-shot generalization across various vision tasks, enabling applications such as image inpainting, outpainting, editing, interpolation, and enrichment, without the need for structural modifications. Our findings suggest that BiGR unifies generative and discriminative tasks effectively, paving the way for further advancements in the field. We further enable BiGR to perform text-to-image generation, showcasing its potential for broader applications.

Xizhe Xue, Yang Zhou, Dawei Yan et al.

Recently, the remarkable success of large language models (LLMs) has achieved a profound impact on the field of artificial intelligence. Numerous advanced works based on LLMs have been proposed and applied in various scenarios. Among them, video language models (VidLMs) are particularly widely used. However, existing works primarily focus on terrestrial scenarios, overlooking the highly demanding application needs of underwater observation. To overcome this gap, we introduce UVLM, an under water observation benchmark which is build through a collaborative approach combining human expertise and AI models. To ensure data quality, we have conducted in-depth considerations from multiple perspectives. First, to address the unique challenges of underwater environments, we selected videos that represent typical underwater challenges including light variations, water turbidity, and diverse viewing angles to construct the dataset. Second, to ensure data diversity, the dataset covers a wide range of frame rates, resolutions, 419 classes of marine animals, and various static plants and terrains. Next, for task diversity, we adopted a structured design where observation targets are categorized into two major classes: biological and environmental. Each category includes content observation and change/action observation, totaling 20 distinct task types. Finally, we designed several challenging evaluation metrics to enable quantitative comparison and analysis of different methods. Experiments on two representative VidLMs demonstrate that fine-tuning VidLMs on UVLM significantly improves underwater world understanding while also showing potential for slight improvements on existing in-air VidLM benchmarks, such as VideoMME and Perception text. The dataset and prompt engineering will be released publicly.

Xianbiao Qi, Yelin He, Jiaquan Ye et al.

Scaling Transformer to a large scale without using some technical tricks such as learning rate warump and using an obviously lower learning rate is an extremely challenging task, and is increasingly gaining more attention. In this paper, we provide a theoretical analysis for the process of training Transformer and reveal the rationale behind the model crash phenomenon in the training process, termed \textit{spectral energy concentration} of ${\bW_q}^{\top} \bW_k$, which is the reason for a malignant entropy collapse, where ${\bW_q}$ and $\bW_k$ are the projection matrices for the query and the key in Transformer, respectively. To remedy this problem, motivated by \textit{Weyl's Inequality}, we present a novel optimization strategy, \ie, making the weight updating in successive steps smooth -- if the ratio $\frac{σ_{1}(\nabla \bW_t)}{σ_{1}(\bW_{t-1})}$ is larger than a threshold, we will automatically bound the learning rate to a weighted multiple of $\frac{σ_{1}(\bW_{t-1})}{σ_{1}(\nabla \bW_t)}$, where $\nabla \bW_t$ is the updating quantity in step $t$. Such an optimization strategy can prevent spectral energy concentration to only a few directions, and thus can avoid malignant entropy collapse which will trigger the model crash. We conduct extensive experiments using ViT, Swin-Transformer and GPT, showing that our optimization strategy can effectively and stably train these Transformers without using learning rate warmup.

Jingyu Xing, Chenwei Tang, Tao Wang et al.

Recent advances in unsupervised anomaly detection (UAD) have shifted from single-class to multi-class scenarios. In such complex contexts, the increasing pattern diversity has brought two challenges to reconstruction-based approaches: (1) over-generalization: anomalies that are subtle or share compositional similarities with normal patterns may be reconstructed with high fidelity, making them difficult to distinguish from normal instances; and (2) insufficient normality reconstruction: complex normal features, such as intricate textures or fine-grained structures, may not be faithfully reconstructed due to the model's limited representational capacity, resulting in false positives. Existing methods typically focus on addressing the former, which unintentionally exacerbate the latter, resulting in inadequate representation of intricate normal patterns. To concurrently address these two challenges, we propose a Memory-augmented Dual-Decoder Networks (MDD-Net). This network includes two critical components: a Dual-Decoder Reverse Distillation Network (DRD-Net) and a Class-aware Memory Module (CMM). Specifically, the DRD-Net incorporates a restoration decoder designed to recover normal features from synthetic abnormal inputs and an identity decoder to reconstruct features that maintain the anomalous semantics. By exploiting the discrepancy between features produced by two decoders, our approach refines anomaly scores beyond the conventional encoder-decoder comparison paradigm, effectively reducing false positives and enhancing localization accuracy. Furthermore, the CMM explicitly encodes and preserves class-specific normal prototypes, actively steering the network away from anomaly reconstruction. Comprehensive experimental results across several benchmarks demonstrate the superior performance of our MDD-Net framework over current SoTA approaches in multi-class UAD tasks.

Penghui Ruan, Bojia Zi, Xianbiao Qi et al.

Object-level manipulation, relocating or reorienting objects in images or videos while preserving scene realism, is central to film post-production, AR, and creative editing. Yet existing methods struggle to jointly achieve three core goals: background preservation, geometric consistency under viewpoint shifts, and user-controllable transformations. Geometry-based approaches offer precise control but require explicit 3D reconstruction and generalize poorly; diffusion-based methods generalize better but lack fine-grained geometric control. We present Ctrl&Shift, an end-to-end diffusion framework to achieve geometry-consistent object manipulation without explicit 3D representations. Our key insight is to decompose manipulation into two stages, object removal and reference-guided inpainting under explicit camera pose control, and encode both within a unified diffusion process. To enable precise, disentangled control, we design a multi-task, multi-stage training strategy that separates background, identity, and pose signals across tasks. To improve generalization, we introduce a scalable real-world dataset construction pipeline that generates paired image and video samples with estimated relative camera poses. Extensive experiments demonstrate that Ctrl&Shift achieves state-of-the-art results in fidelity, viewpoint consistency, and controllability. To our knowledge, this is the first framework to unify fine-grained geometric control and real-world generalization for object manipulation, without relying on any explicit 3D modeling.

Dongquan Yang, Yifan Yang, Xiaotian Yu et al.

Processing long-context inputs with large language models presents a significant challenge due to the enormous memory requirements of the Key-Value (KV) cache during inference. Existing KV cache compression methods exhibit noticeable performance degradation when memory is reduced by more than 85%. Additionally, strategies that leverage GPU-CPU collaboration for approximate attention remain underexplored in this setting. We propose HCAttention, a heterogeneous attention computation framework that integrates key quantization, value offloading, and dynamic KV eviction to enable efficient inference under extreme memory constraints. The method is compatible with existing transformer architectures and does not require model fine-tuning. Experimental results on the LongBench benchmark demonstrate that our approach preserves the accuracy of full-attention model while shrinking the KV cache memory footprint to 25% of its original size. Remarkably, it stays competitive with only 12.5% of the cache, setting a new state-of-the-art in LLM KV cache compression. To the best of our knowledge, HCAttention is the first to extend the Llama-3-8B model to process 4 million tokens on a single A100 GPU with 80GB memory.

Feng Lin, Marco Chen, Haokui Zhang et al.

This paper investigates the role of attention heads in CLIP's image encoder. Building on interpretability studies, we conduct an exhaustive analysis and find that certain heads, distributed across layers, are detrimental to the resulting representations. To mitigate their impact, we propose a simple yet effective Attention Ablation Technique (AAT) that suppresses selected heads by directly manipulating their attention weights. By incorporating two complementary strategies tailored to different application scenarios, AAT enables the systematic identification and ablation of harmful heads with minimal overhead. Experiments show that AAT consistently improves downstream performance across diverse domains, boosting recall by up to 11.1% on cross-modal retrieval benchmarks. These results highlight that AAT can effectively refine large-scale VLMs with virtually no extra inference cost, while yielding semantically meaningful patterns that align with existing interpretability findings.

Haizhao Jing, Haokui Zhang, Zhenhao Shang et al.

Neural Architecture Representation Learning aims to transform network models into feature representations for predicting network attributes, playing a crucial role in deploying and designing networks for real-world applications. Recently, inspired by the success of transformers, transformer-based models integrated with Graph Neural Networks (GNNs) have achieved significant progress in representation learning. However, current methods still have some limitations. First, existing methods overlook hardware attribute information, which conflicts with the current trend of diversified deep learning hardware and limits the practical applicability of models. Second, current encoding approaches rely on static adjacency matrices to represent topological structures, failing to capture the structural differences between computational nodes, which ultimately compromises encoding effectiveness. In this paper, we introduce LeDG-Former, an innovative framework that addresses these limitations through the synergistic integration of language-based semantic embedding and dynamic graph representation learning. Specifically, inspired by large language models (LLMs), we propose a language embedding framework where both neural architectures and hardware platform specifications are projected into a unified semantic space through tokenization and LLM processing, enabling zero-shot prediction across different hardware platforms for the first time. Then, we propose a dynamic graph-based transformer for modeling neural architectures, resulting in improved neural architecture modeling performance. On the NNLQP benchmark, LeDG-Former surpasses previous methods, establishing a new SOTA while demonstrating the first successful cross-hardware latency prediction capability. Furthermore, our framework achieves superior performance on the cell-structured NAS-Bench-101 and NAS-Bench-201 datasets.

Wenqiang Zhou, Zhendong Yu, Xinyu Liu et al.

Quantization Aware Training (QAT) is a neural network quantization technique that compresses model size and improves operational efficiency while effectively maintaining model performance. The paradigm of QAT is to introduce fake quantization operators during the training process, allowing the model to autonomously compensate for information loss caused by quantization. Making quantization parameters trainable can significantly improve the performance of QAT, but at the cost of compromising the flexibility during inference, especially when dealing with activation values with substantially different distributions. In this paper, we propose an effective learnable adaptive neural network quantization method, called Adaptive Step Size Quantization (ASQ), to resolve this conflict. Specifically, the proposed ASQ method first dynamically adjusts quantization scaling factors through a trained module capable of accommodating different activations. Then, to address the rigid resolution issue inherent in Power of Two (POT) quantization, we propose an efficient non-uniform quantization scheme. We utilize the Power Of Square root of Two (POST) as the basis for exponential quantization, effectively handling the bell-shaped distribution of neural network weights across various bit-widths while maintaining computational efficiency through a Look-Up Table method (LUT). Extensive experimental results demonstrate that the proposed ASQ method is superior to the state-of-the-art QAT approaches. Notably that the ASQ is even competitive compared to full precision baselines, with its 4-bit quantized ResNet34 model improving accuracy by 1.2\% on ImageNet.

Yuxiang Zhong, Xianbiao Qi, Shanjun Li et al.

In this technical report, we present our 1st place solution for the ICDAR 2021 competition on mathematical formula detection (MFD). The MFD task has three key challenges including a large scale span, large variation of the ratio between height and width, and rich character set and mathematical expressions. Considering these challenges, we used Generalized Focal Loss (GFL), an anchor-free method, instead of the anchor-based method, and prove the Adaptive Training Sampling Strategy (ATSS) and proper Feature Pyramid Network (FPN) can well solve the important issue of scale variation. Meanwhile, we also found some tricks, e.g., Deformable Convolution Network (DCN), SyncBN, and Weighted Box Fusion (WBF), were effective in MFD task. Our proposed method ranked 1st in the final 15 teams.

Yong Qian, Zhongqing Wang, Rong Xiao et al.

Previous studies show effective of pre-trained language models for sentiment analysis. However, most of these studies ignore the importance of sentimental information for pre-trained models.Therefore, we fully investigate the sentimental information for pre-trained models and enhance pre-trained language models with semantic graphs for sentiment analysis.In particular, we introduce Semantic Graphs based Pre-training(SGPT) using semantic graphs to obtain synonym knowledge for aspect-sentiment pairs and similar aspect/sentiment terms.We then optimize the pre-trained language model with the semantic graphs.Empirical studies on several downstream tasks show that proposed model outperforms strong pre-trained baselines. The results also show the effectiveness of proposed semantic graphs for pre-trained model.

Houyi Li, Zhihong Chen, Chenliang Li et al.

The large-scale recommender system mainly consists of two stages: matching and ranking. The matching stage (also known as the retrieval step) identifies a small fraction of relevant items from billion-scale item corpus in low latency and computational cost. Item-to-item collaborative filter (item-based CF) and embedding-based retrieval (EBR) have been long used in the industrial matching stage owing to its efficiency. However, item-based CF is hard to meet personalization, while EBR has difficulty in satisfying diversity. In this paper, we propose a novel matching architecture, Path-based Deep Network (named PDN), which can incorporate both personalization and diversity to enhance matching performance. Specifically, PDN is comprised of two modules: Trigger Net and Similarity Net. PDN utilizes Trigger Net to capture the user's interest in each of his/her interacted item, and Similarity Net to evaluate the similarity between each interacted item and the target item based on these items' profile and CF information. The final relevance between the user and the target item is calculated by explicitly considering user's diverse interests, \ie aggregating the relevance weights of the related two-hop paths (one hop of a path corresponds to user-item interaction and the other to item-item relevance). Furthermore, we describe the architecture design of a matching system with the proposed PDN in a leading real-world E-Commerce service (Mobile Taobao App). Based on offline evaluations and online A/B test, we show that PDN outperforms the existing solutions for the same task. The online results also demonstrate that PDN can retrieve more personalized and more diverse relevant items to significantly improve user engagement. Currently, PDN system has been successfully deployed at Mobile Taobao App and handling major online traffic.

Jiaquan Ye, Xianbiao Qi, Yelin He et al.

This paper presents our solution for ICDAR 2021 competition on scientific literature parsing taskB: table recognition to HTML. In our method, we divide the table content recognition task into foursub-tasks: table structure recognition, text line detection, text line recognition, and box assignment.Our table structure recognition algorithm is customized based on MASTER [1], a robust image textrecognition algorithm. PSENet [2] is used to detect each text line in the table image. For text linerecognition, our model is also built on MASTER. Finally, in the box assignment phase, we associatedthe text boxes detected by PSENet with the structure item reconstructed by table structure prediction,and fill the recognized content of the text line into the corresponding item. Our proposed methodachieves a 96.84% TEDS score on 9,115 validation samples in the development phase, and a 96.32%TEDS score on 9,064 samples in the final evaluation phase.

Yelin He, Xianbiao Qi, Jiaquan Ye et al.

This paper presents our solution for the ICDAR 2021 Competition on Scientific Table Image Recognition to LaTeX. This competition has two sub-tasks: Table Structure Reconstruction (TSR) and Table Content Reconstruction (TCR). We treat both sub-tasks as two individual image-to-sequence recognition problems. We leverage our previously proposed algorithm MASTER \cite{lu2019master}, which is originally proposed for scene text recognition. We optimize the MASTER model from several perspectives: network structure, optimizer, normalization method, pre-trained model, resolution of input image, data augmentation, and model ensemble. Our method achieves 0.7444 Exact Match and 0.8765 Exact Match @95\% on the TSR task, and obtains 0.5586 Exact Match and 0.7386 Exact Match 95\% on the TCR task.

Shaowei Yao, Jiwei Tan, Xi Chen et al.

The search engine plays a fundamental role in online e-commerce systems, to help users find the products they want from the massive product collections. Relevance is an essential requirement for e-commerce search, since showing products that do not match search query intent will degrade user experience. With the existence of vocabulary gap between user language of queries and seller language of products, measuring semantic relevance is necessary and neural networks are engaged to address this task. However, semantic relevance is different from click-through rate prediction in that no direct training signal is available. Most previous attempts learn relevance models from user click-through data that are cheap and abundant. Unfortunately, click behavior is noisy and misleading, which is affected by not only relevance but also factors including price, image and attractive titles. Therefore, it is challenging but valuable to learn relevance models from click-through data. In this paper, we propose a new relevance learning framework that concentrates on how to train a relevance model from the weak supervision of click-through data. Different from previous efforts that treat samples as either relevant or irrelevant, we construct more fine-grained samples for training. We propose a novel way to consider samples of different relevance confidence, and come up with a new training objective to learn a robust relevance model with desirable score distribution. The proposed model is evaluated on offline annotated data and online A/B testing, and it achieves both promising performance and high computational efficiency. The model has already been deployed online, serving the search traffic of Taobao for over a year.

Bingcong Li, Xin Tang, Xianbiao Qi et al.

Recently, inspired by Transformer, self-attention-based scene text recognition approaches have achieved outstanding performance. However, we find that the size of model expands rapidly with the lexicon increasing. Specifically, the number of parameters for softmax classification layer and output embedding layer are proportional to the vocabulary size. It hinders the development of a lightweight text recognition model especially applied for Chinese and multiple languages. Thus, we propose a lightweight scene text recognition model named Hamming OCR. In this model, a novel Hamming classifier, which adopts locality sensitive hashing (LSH) algorithm to encode each character, is proposed to replace the softmax regression and the generated LSH code is directly employed to replace the output embedding. We also present a simplified transformer decoder to reduce the number of parameters by removing the feed-forward network and using cross-layer parameter sharing technique. Compared with traditional methods, the number of parameters in both classification and embedding layers is independent on the size of vocabulary, which significantly reduces the storage requirement without loss of accuracy. Experimental results on several datasets, including four public benchmaks and a Chinese text dataset synthesized by SynthText with more than 20,000 characters, shows that Hamming OCR achieves competitive results.

Cheng Zhao, Chenliang Li, Rong Xiao et al.

In a large recommender system, the products (or items) could be in many different categories or domains. Given two relevant domains (e.g., Book and Movie), users may have interactions with items in one domain but not in the other domain. To the latter, these users are considered as cold-start users. How to effectively transfer users' preferences based on their interactions from one domain to the other relevant domain, is the key issue in cross-domain recommendation. Inspired by the advances made in review-based recommendation, we propose to model user preference transfer at aspect-level derived from reviews. To this end, we propose a cross-domain recommendation framework via aspect transfer network for cold-start users (named CATN). CATN is devised to extract multiple aspects for each user and each item from their review documents, and learn aspect correlations across domains with an attention mechanism. In addition, we further exploit auxiliary reviews from like-minded users to enhance a user's aspect representations. Then, an end-to-end optimization framework is utilized to strengthen the robustness of our model. On real-world datasets, the proposed CATN outperforms SOTA models significantly in terms of rating prediction accuracy. Further analysis shows that our model is able to reveal user aspect connections across domains at a fine level of granularity, making the recommendation explainable.

Di Wu, Yihao Chen, Xianbiao Qi et al.

How can we effectively utilise the 2D monocular image information for recovering the 6D pose (6-DoF) of the visual objects? Deep learning has shown to be effective for robust and real-time monocular pose estimation. Oftentimes, the network learns to regress the 6-DoF pose using a naive loss function. However, due to a lack of geometrical scene understanding from the directly regressed pose estimation, there are misalignments between the rendered mesh from the 3D object and the 2D instance segmentation result, e.g., bounding boxes and masks prediction. This paper bridges the gap between 2D mask generation and 3D location prediction via a differentiable neural mesh renderer. We utilise the overlay between the accurate mask prediction and less accurate mesh prediction to iteratively optimise the direct regressed 6D pose information with a focus on translation estimation. By leveraging geometry, we demonstrate that our technique significantly improves direct regression performance on the difficult task of translation estimation and achieve the state of the art results on Peking University/Baidu - Autonomous Driving dataset and the ApolloScape 3D Car Instance dataset. The code can be found at \url{https://bit.ly/2IRihfU}.

Jianfeng Wang, Rong Xiao, Yandong Guo et al.

In this paper, we study the problem of object counting with incomplete annotations. Based on the observation that in many object counting problems the target objects are normally repeated and highly similar to each other, we are particularly interested in the setting when only a few exemplar annotations are provided. Directly applying object detection with incomplete annotations will result in severe accuracy degradation due to its improper handling of unlabeled object instances. To address the problem, we propose a positiveness-focused object detector (PFOD) to progressively propagate the incomplete labels before applying the general object detection algorithm. The PFOD focuses on the positive samples and ignore the negative instances at most of the learning time. This strategy, though simple, dramatically boosts the object counting accuracy. On the CARPK dataset for parking lot car counting, we improved mAP@0.5 from 4.58% to 72.44% using only 5 training images each with 5 bounding boxes. On the Drink35 dataset for shelf product counting, the mAP@0.5 is improved from 14.16% to 53.73% using 10 training images each with 5 bounding boxes.

Bowen Cheng, Rong Xiao, Jianfeng Wang et al.

We present a novel high frequency residual learning framework, which leads to a highly efficient multi-scale network (MSNet) architecture for mobile and embedded vision problems. The architecture utilizes two networks: a low resolution network to efficiently approximate low frequency components and a high resolution network to learn high frequency residuals by reusing the upsampled low resolution features. With a classifier calibration module, MSNet can dynamically allocate computation resources during inference to achieve a better speed and accuracy trade-off. We evaluate our methods on the challenging ImageNet-1k dataset and observe consistent improvements over different base networks. On ResNet-18 and MobileNet with alpha=1.0, MSNet gains 1.5% accuracy over both architectures without increasing computations. On the more efficient MobileNet with alpha=0.25, our method gains 3.8% accuracy with the same amount of computations.

Bowen Cheng, Rong Xiao, Yandong Guo et al.

We study in this paper how to initialize the parameters of multinomial logistic regression (a fully connected layer followed with softmax and cross entropy loss), which is widely used in deep neural network (DNN) models for classification problems. As logistic regression is widely known not having a closed-form solution, it is usually randomly initialized, leading to several deficiencies especially in transfer learning where all the layers except for the last task-specific layer are initialized using a pre-trained model. The deficiencies include slow convergence speed, possibility of stuck in local minimum, and the risk of over-fitting. To address those deficiencies, we first study the properties of logistic regression and propose a closed-form approximate solution named regularized Gaussian classifier (RGC). Then we adopt this approximate solution to initialize the task-specific linear layer and demonstrate superior performance over random initialization in terms of both accuracy and convergence speed on various tasks and datasets. For example, for image classification, our approach can reduce the training time by 10 times and achieve 3.2% gain in accuracy for Flickr-style classification. For object detection, our approach can also be 10 times faster in training for the same accuracy, or 5% better in terms of mAP for VOC 2007 with slightly longer training.