Weijie Zhao

Ping Li, Weijie Zhao

This report presents the open-source package which implements the series of our boosting works in the past years. In particular, the package includes mainly three lines of techniques, among which the following two are already the standard implementations in popular boosted tree platforms: (i) The histogram-based (feature-binning) approach makes the tree implementation convenient and efficient. In Li et al (2007), a simple fixed-length adaptive binning algorithm was developed. In this report, we demonstrate that such a simple algorithm is still surprisingly effective compared to more sophisticated variants in popular tree platforms. (ii) The explicit gain formula in Li (20010) for tree splitting based on second-order derivatives of the loss function typically improves, often considerably, over the first-order methods. Although the gain formula in Li (2010) was derived for logistic regression loss, it is a generic formula for loss functions with second-derivatives. For example, the open-source package also includes $L_p$ regression for $p\geq 1$. The main contribution of this package is the ABC-Boost (adaptive base class boosting) for multi-class classification. The initial work in Li (2008) derived a new set of derivatives of the classical multi-class logistic regression by specifying a "base class". The accuracy can be substantially improved if the base class is chosen properly. The major technical challenge is to design a search strategy to select the base class. The prior published works implemented an exhaustive search procedure to find the base class which is computationally too expensive. Recently, a new report (Li and Zhao, 20022) presents a unified framework of "Fast ABC-Boost" which allows users to efficiently choose the proper search space for the base class. The package provides interfaces for linux, windows, mac, matlab, R, python.

Ping Li, Weijie Zhao

The work in ICML'09 showed that the derivatives of the classical multi-class logistic regression loss function could be re-written in terms of a pre-chosen "base class" and applied the new derivatives in the popular boosting framework. In order to make use of the new derivatives, one must have a strategy to identify/choose the base class at each boosting iteration. The idea of "adaptive base class boost" (ABC-Boost) in ICML'09, adopted a computationally expensive "exhaustive search" strategy for the base class at each iteration. It has been well demonstrated that ABC-Boost, when integrated with trees, can achieve substantial improvements in many multi-class classification tasks. Furthermore, the work in UAI'10 derived the explicit second-order tree split gain formula which typically improved the classification accuracy considerably, compared with using only the fist-order information for tree-splitting, for both multi-class and binary-class classification tasks. In this paper, we develop a unified framework for effectively selecting the base class by introducing a series of ideas to improve the computational efficiency of ABC-Boost. Our framework has parameters $(s,g,w)$. At each boosting iteration, we only search for the "$s$-worst classes" (instead of all classes) to determine the base class. We also allow a "gap" $g$ when conducting the search. That is, we only search for the base class at every $g+1$ iterations. We furthermore allow a "warm up" stage by only starting the search after $w$ boosting iterations. The parameters $s$, $g$, $w$, can be viewed as tunable parameters and certain combinations of $(s,g,w)$ may even lead to better test accuracy than the "exhaustive search" strategy. Overall, our proposed framework provides a robust and reliable scheme for implementing ABC-Boost in practice.

Ping Li, Weijie Zhao

In this work, we demonstrate the advantage of the pGMM (``powered generalized min-max'') kernel in the context of (ridge) regression. In recent prior studies, the pGMM kernel has been extensively evaluated for classification tasks, for logistic regression, support vector machines, as well as deep neural networks. In this paper, we provide an experimental study on ridge regression, to compare the pGMM kernel regression with the ordinary ridge linear regression as well as the RBF kernel ridge regression. Perhaps surprisingly, even without a tuning parameter (i.e., $p=1$ for the power parameter of the pGMM kernel), the pGMM kernel already performs well. Furthermore, by tuning the parameter $p$, this (deceptively simple) pGMM kernel even performs quite comparably to boosted trees. Boosting and boosted trees are very popular in machine learning practice. For regression tasks, typically, practitioners use $L_2$ boost, i.e., for minimizing the $L_2$ loss. Sometimes for the purpose of robustness, the $L_1$ boost might be a choice. In this study, we implement $L_p$ boost for $p\geq 1$ and include it in the package of ``Fast ABC-Boost''. Perhaps also surprisingly, the best performance (in terms of $L_2$ regression loss) is often attained at $p>2$, in some cases at $p\gg 2$. This phenomenon has already been demonstrated by Li et al (UAI 2010) in the context of k-nearest neighbor classification using $L_p$ distances. In summary, the implementation of $L_p$ boost provides practitioners the additional flexibility of tuning boosting algorithms for potentially achieving better accuracy in regression applications.

Weijie Zhao, Xuewu Jiao, Xinsheng Luo et al.

Deep learning has been widely deployed for online ads systems to predict Click-Through Rate (CTR). Machine learning researchers and practitioners frequently retrain CTR models to test their new extracted features. However, the CTR model training often relies on a large number of raw input data logs. Hence, the feature extraction can take a significant proportion of the training time for an industrial-level CTR model. In this paper, we propose FeatureBox, a novel end-to-end training framework that pipelines the feature extraction and the training on GPU servers to save the intermediate I/O of the feature extraction. We rewrite computation-intensive feature extraction operators as GPU operators and leave the memory-intensive operator on CPUs. We introduce a layer-wise operator scheduling algorithm to schedule these heterogeneous operators. We present a light-weight GPU memory management algorithm that supports dynamic GPU memory allocation with minimal overhead. We experimentally evaluate FeatureBox and compare it with the previous in-production feature extraction framework on two real-world ads applications. The results confirm the effectiveness of our proposed method.

Weijie Zhao, Huajie Shao, Zhaozhuo Xu et al.

Exploring the data sources used to train Large Language Models (LLMs) is a crucial direction in investigating potential copyright infringement by these models. While this approach can identify the possible use of copyrighted materials in training data, it does not directly measure infringing risks. Recent research has shifted towards testing whether LLMs can directly output copyrighted content. Addressing this direction, we investigate and assess LLMs' capacity to generate infringing content by providing them with partial information from copyrighted materials, and try to use iterative prompting to get LLMs to generate more infringing content. Specifically, we input a portion of a copyrighted text into LLMs, prompt them to complete it, and then analyze the overlap between the generated content and the original copyrighted material. Our findings demonstrate that LLMs can indeed generate content highly overlapping with copyrighted materials based on these partial inputs.

Jun Woo Chung, Yingjie Lao, Weijie Zhao

Gradient Boosting Decision Trees (GBDTs) are widely used in industry and academia for their high accuracy and efficiency, particularly on structured data. However, watermarking GBDT models remains underexplored compared to neural networks. In this work, we present the first robust watermarking framework tailored to GBDT models, utilizing in-place fine-tuning to embed imperceptible and resilient watermarks. We propose four embedding strategies, each designed to minimize impact on model accuracy while ensuring watermark robustness. Through experiments across diverse datasets, we demonstrate that our methods achieve high watermark embedding rates, low accuracy degradation, and strong resistance to post-deployment fine-tuning.

Weijie Zhao, Ping Li

Feature interactions can play a crucial role in recommendation systems as they capture complex relationships between user preferences and item characteristics. Existing methods such as Deep & Cross Network (DCNv2) may suffer from high computational requirements due to their cross-layer operations. In this paper, we propose a novel approach called blockwise feature interaction (BFI) to help alleviate this issue. By partitioning the feature interaction process into smaller blocks, we can significantly reduce both the memory footprint and the computational burden. Four variants (denoted by P, Q, T, S, respectively) of BFI have been developed and empirically compared. Our experimental results demonstrate that the proposed algorithms achieves close accuracy compared to the standard DCNv2, while greatly reducing the computational overhead and the number of parameters. This paper contributes to the development of efficient recommendation systems by providing a practical solution for improving feature interaction efficiency.

Ping Li, Weijie Zhao

Many hashing algorithms including minwise hashing (MinHash), one permutation hashing (OPH), and consistent weighted sampling (CWS) generate integers of $B$ bits. With $k$ hashes for each data vector, the storage would be $B\times k$ bits; and when used for large-scale learning, the model size would be $2^B\times k$, which can be expensive. A standard strategy is to use only the lowest $b$ bits out of the $B$ bits and somewhat increase $k$, the number of hashes. In this study, we propose to re-use the hashes by partitioning the $B$ bits into $m$ chunks, e.g., $b\times m =B$. Correspondingly, the model size becomes $m\times 2^b \times k$, which can be substantially smaller than the original $2^B\times k$. Our theoretical analysis reveals that by partitioning the hash values into $m$ chunks, the accuracy would drop. In other words, using $m$ chunks of $B/m$ bits would not be as accurate as directly using $B$ bits. This is due to the correlation from re-using the same hash. On the other hand, our analysis also shows that the accuracy would not drop much for (e.g.,) $m=2\sim 4$. In some regions, Pb-Hash still works well even for $m$ much larger than 4. We expect Pb-Hash would be a good addition to the family of hashing methods/applications and benefit industrial practitioners. We verify the effectiveness of Pb-Hash in machine learning tasks, for linear SVM models as well as deep learning models. Since the hashed data are essentially categorical (ID) features, we follow the standard practice of using embedding tables for each hash. With Pb-Hash, we need to design an effective strategy to combine $m$ embeddings. Our study provides an empirical evaluation on four pooling schemes: concatenation, max pooling, mean pooling, and product pooling. There is no definite answer which pooling would be always better and we leave that for future study.

Ziwen Liu, Huawei Lin, Yide Ran et al.

Large language models (LLMs) sometimes memorize undesirable knowledge, which must be removed after deployment. Prior work on machine unlearning has focused largely on optimization methods that adjust parameters to enforce forgetting while preserving retention. However, these approaches assume that the forget and retain sets are readily available, which rarely holds in practice. Unlearning is typically triggered by an undesired generation at inference time, making the retrieval of relevant data the central challenge. We introduce the notion of data Pareto improvement for LLM unlearning, which formalizes how retrieval can expand the achievable trade-off frontier between forgetting and retention. To realize this principle, we propose Randomized Antipodal Search on Linearized Influence Kernel (RASLIK), a retrieval algorithm that combines permutation-projection hashing with randomized antipodal search. RASLIK reduces selection variance, achieves sublinear complexity, and yields a double gain in both quality and efficiency. Across multiple models, datasets, and unlearning algorithms, RASLIK consistently outperforms deterministic baselines and even oracle sampling, establishing randomized search as a principled and scalable solution for data-centric unlearning.

Weijie Zhao, Mingquan Liu, Bolun Wang et al.

Scaling Transformers typically necessitates training larger models from scratch, as standard architectures struggle to expand without discarding learned representations. We identify the primary bottleneck in the attention mechanism's linear projections, which strictly confine feature extraction to fixed-dimensional subspaces, limiting both expressivity and incremental capacity. To address this, we introduce Nexusformer, which replaces linear $Q/K/V$ projections with a Nexus-Rank layer, a three-stage nonlinear mapping driven by dual activations in progressively higher dimensional spaces. This design overcomes the linearity constraint and enables lossless structured growth: new capacity can be injected along two axes via zero-initialized blocks that preserve pretrained knowledge. Experiments on language modeling and reasoning benchmarks demonstrate that Nexusformer matches Tokenformer's perplexity using up to 41.5\% less training compute during progressive scaling (240M to 440M). Furthermore, our analysis of growth dynamics reveals that zero initialization induces a stable convergence trajectory, allowing us to derive a geometric scaling law that accurately predicts performance across expansion scales.

Rui Chu, Bingyin Zhao, Thanh Quoc Hung Le et al.

Large language models (LLMs) have achieved unprecedented success due to their exceptional generative capabilities. However, because they depend on knowledge encapsulated from training corpora, they may produce hallucinations, stereotypes, and socially biased content. In particular, LLMs are prone to prejudiced responses involving race, gender, and age, which are collectively referred to as social biases. Prior studies have used fine-tuning and prompt engineering to mitigate such biases in LLMs, but these methods require additional training resources or domain knowledge to design the framework. Moreover, they may degrade the original capabilities of LLMs and often overlook the need for dynamic debiasing contexts for fairer inference. In this paper, we propose DebiasRAG, a novel tuning-free and dynamic query-specific debiasing framework based on retrieval-augmented generation (RAG). DebiasRAG improves fairness while preserving the intrinsic properties of LLMs, such as representation ability. DebiasRAG consists of three stages: (1) query-specific debiasing candidate generation; (2) context candidate pool construction; and (3) gradient-updated debiasing-guided context piece reranking. First, DebiasRAG leverages self-diagnosed bias contexts relevant to the query through regular retrieval, where the bias contexts are prepared offline by the DebiasRAG provider. Given the query-specific bias contexts, DebiasRAG reversely produces debiasing contexts, which are provided as additional fairness constraints for LLM outputs. Second, a regular RAG retrieval process produces query-related contexts from the regular RAG document database, such as a chunked Wikipedia dataset.

Huawei Lin, Tong Geng, Zhaozhuo Xu et al.

Autoregressive (AR) models have recently shown strong performance in image generation, where a critical component is the visual tokenizer (VT) that maps continuous pixel inputs to discrete token sequences. The quality of the VT largely defines the upper bound of AR model performance. However, current discrete VTs fall significantly behind continuous variational autoencoders (VAEs), leading to degraded image reconstructions and poor preservation of details and text. Existing benchmarks focus on end-to-end generation quality, without isolating VT performance. To address this gap, we introduce VTBench, a comprehensive benchmark that systematically evaluates VTs across three core tasks: Image Reconstruction, Detail Preservation, and Text Preservation, and covers a diverse range of evaluation scenarios. We systematically assess state-of-the-art VTs using a set of metrics to evaluate the quality of reconstructed images. Our findings reveal that continuous VAEs produce superior visual representations compared to discrete VTs, particularly in retaining spatial structure and semantic detail. In contrast, the degraded representations produced by discrete VTs often lead to distorted reconstructions, loss of fine-grained textures, and failures in preserving text and object integrity. Furthermore, we conduct experiments on GPT-4o image generation and discuss its potential AR nature, offering new insights into the role of visual tokenization. We release our benchmark and codebase publicly to support further research and call on the community to develop strong, general-purpose open-source VTs.

Guoshenghui Zhao, Huawei Lin, Weijie Zhao

Removing specific data influence from large language models (LLMs) remains challenging, as retraining is costly and existing approximate unlearning methods are often unstable. The challenge is exacerbated when the forget set is small or imbalanced. We introduce RapidUn, an influence-driven and parameter-efficient unlearning framework. It first estimates per-sample influence through a fast estimation module, then maps these scores into adaptive update weights that guide selective parameter updates -- forgetting harmful behavior while retaining general knowledge. On Mistral-7B and Llama-3-8B across Dolly-15k and Alpaca-57k, RapidUn achieves up to 100 times higher efficiency than full retraining and consistently outperforms Fisher, GA, and LoReUn on both in-distribution and out-of-distribution forgetting. These results establish influence-guided parameter reweighting as a scalable and interpretable paradigm for LLM unlearning.

Huawei Lin, Yunzhi Shi, Tong Geng et al.

Multimodal large language models (MLLMs) have shown strong capabilities but remain limited to fixed modality pairs and require costly fine-tuning with large aligned datasets. Building fully omni-capable models that can integrate text, images, audio, and video remains impractical and lacks robust reasoning support. In this paper, we propose an Agent-Omni framework that coordinates existing foundation models through a master-agent system, enabling flexible multimodal reasoning without retraining. The master agent interprets user intent, delegates subtasks to modality-specific agents, and integrates their outputs into coherent responses. Extensive experiments across text, image, audio, video, and omni benchmarks show that Agent-Omni consistently achieves state-of-the-art performance, particularly on tasks requiring complex cross-modal reasoning. Its agent-based design enables seamless integration of specialized foundation models, ensuring adaptability to diverse inputs while maintaining transparency and interpretability. In addition, the framework is modular and easily extensible, allowing future improvements as stronger models become available.

Guoshenghui Zhao, Weijie Zhao, Tan Yu

Diffusion large language models generate text through multi-step denoising, where hallucination signals may emerge throughout the trajectory rather than only in the final output. Existing detectors mainly rely on output uncertainty or coarse trace statistics, which often fail to capture the richer hidden dynamics of D-LLMs. We propose HIVE, a hidden-evidence verification framework that extracts compressed hidden evidence from denoising trajectories, selects informative step-layer evidence, and conditions a verifier language model on the selected evidence through prefix embeddings. HIVE produces both a continuous hallucination score from verifier decision logits and structured verification outputs, including hallucination types, evidence pairs, and short rationales. Across two D-LLMs and three QA benchmarks, HIVE consistently outperforms eight strong baselines and achieves up to 0.9236 AUROC and 0.9537 AUPRC. Ablation studies further confirm the importance of hidden-evidence conditioning, learned evidence selection, two-stream evidence representation, and step-layer embeddings. These results suggest that selected hidden evidence from denoising trajectories provides a stronger and more usable hallucination signal than output-only uncertainty or coarse trace statistics.

Huawei Lin, Jikai Long, Zhaozhuo Xu et al.

Given a Large Language Model (LLM) generation, how can we identify which training data led to this generation? In this paper, we proposed RapidIn, a scalable framework adapting to LLMs for estimating the influence of each training data. The proposed framework consists of two stages: caching and retrieval. First, we compress the gradient vectors by over 200,000x, allowing them to be cached on disk or in GPU/CPU memory. Then, given a generation, RapidIn efficiently traverses the cached gradients to estimate the influence within minutes, achieving over a 6,326x speedup. Moreover, RapidIn supports multi-GPU parallelization to substantially accelerate caching and retrieval. Our empirical result confirms the efficiency and effectiveness of RapidIn.

Huawei Lin, Yingjie Lao, Tong Geng et al.

Large Language Models (LLMs) are vulnerable to attacks like prompt injection, backdoor attacks, and adversarial attacks, which manipulate prompts or models to generate harmful outputs. In this paper, departing from traditional deep learning attack paradigms, we explore their intrinsic relationship and collectively term them Prompt Trigger Attacks (PTA). This raises a key question: Can we determine if a prompt is benign or poisoned? To address this, we propose UniGuardian, the first unified defense mechanism designed to detect prompt injection, backdoor attacks, and adversarial attacks in LLMs. Additionally, we introduce a single-forward strategy to optimize the detection pipeline, enabling simultaneous attack detection and text generation within a single forward pass. Our experiments confirm that UniGuardian accurately and efficiently identifies malicious prompts in LLMs.

Huawei Lin, Yingjie Lao, Weijie Zhao

Identifying the training data samples that most influence a generated image is a critical task in understanding diffusion models (DMs), yet existing influence estimation methods are constrained to small-scale or LoRA-tuned models due to computational limitations. To address this challenge, we propose DMin (Diffusion Model influence), a scalable framework for estimating the influence of each training data sample on a given generated image. To the best of our knowledge, it is the first method capable of influence estimation for DMs with billions of parameters. Leveraging efficient gradient compression, DMin reduces storage requirements from hundreds of TBs to mere MBs or even KBs, and retrieves the top-k most influential training samples in under 1 second, all while maintaining performance. Our empirical results demonstrate DMin is both effective in identifying influential training samples and efficient in terms of computational and storage requirements.

Yanzhou Pan, Huawei Lin, Yide Ran et al.

Large Language Models (LLMs) heavily rely on high-quality training data, making data valuation crucial for optimizing model performance, especially when working within a limited budget. In this work, we aim to offer a third-party data valuation approach that benefits both data providers and model developers. We introduce a linearized future influence kernel (LinFiK), which assesses the value of individual data samples in improving LLM performance during training. We further propose ALinFiK, a learning strategy to approximate LinFiK, enabling scalable data valuation. Our comprehensive evaluations demonstrate that this approach surpasses existing baselines in effectiveness and efficiency, demonstrating significant scalability advantages as LLM parameters increase.

Huawei Lin, Jun Woo Chung, Yingjie Lao et al.

Gradient Boosting Decision Tree (GBDT) is one of the most popular machine learning models in various applications. However, in the traditional settings, all data should be simultaneously accessed in the training procedure: it does not allow to add or delete any data instances after training. In this paper, we propose an efficient online learning framework for GBDT supporting both incremental and decremental learning. To the best of our knowledge, this is the first work that considers an in-place unified incremental and decremental learning on GBDT. To reduce the learning cost, we present a collection of optimizations for our framework, so that it can add or delete a small fraction of data on the fly. We theoretically show the relationship between the hyper-parameters of the proposed optimizations, which enables trading off accuracy and cost on incremental and decremental learning. The backdoor attack results show that our framework can successfully inject and remove backdoor in a well-trained model using incremental and decremental learning, and the empirical results on public datasets confirm the effectiveness and efficiency of our proposed online learning framework and optimizations.

Duy C. Hoang, Hung T. Q. Le, Rui Chu et al.

With the widespread adoption of Large Language Models (LLMs), concerns about potential misuse have emerged. To this end, watermarking has been adapted to LLM, enabling a simple and effective way to detect and monitor generated text. However, while the existing methods can differentiate between watermarked and unwatermarked text with high accuracy, they often face a trade-off between the quality of the generated text and the effectiveness of the watermarking process. In this work, we present a novel type of LLM watermark, Sparse Watermark, which aims to mitigate this trade-off by applying watermarks to a small subset of generated tokens distributed across the text. The key strategy involves anchoring watermarked tokens to words that have specific Part-of-Speech (POS) tags. Our experimental results demonstrate that the proposed watermarking scheme achieves high detectability while generating text that outperforms previous LLM watermarking methods in quality across various tasks

Ping Li, Weijie Zhao

We develop the "generalized consistent weighted sampling" (GCWS) for hashing the "powered-GMM" (pGMM) kernel (with a tuning parameter $p$). It turns out that GCWS provides a numerically stable scheme for applying power transformation on the original data, regardless of the magnitude of $p$ and the data. The power transformation is often effective for boosting the performance, in many cases considerably so. We feed the hashed data to neural networks on a variety of public classification datasets and name our method ``GCWSNet''. Our extensive experiments show that GCWSNet often improves the classification accuracy. Furthermore, it is evident from the experiments that GCWSNet converges substantially faster. In fact, GCWS often reaches a reasonable accuracy with merely (less than) one epoch of the training process. This property is much desired because many applications, such as advertisement click-through rate (CTR) prediction models, or data streams (i.e., data seen only once), often train just one epoch. Another beneficial side effect is that the computations of the first layer of the neural networks become additions instead of multiplications because the input data become binary (and highly sparse). Empirical comparisons with (normalized) random Fourier features (NRFF) are provided. We also propose to reduce the model size of GCWSNet by count-sketch and develop the theory for analyzing the impact of using count-sketch on the accuracy of GCWS. Our analysis shows that an ``8-bit'' strategy should work well in that we can always apply an 8-bit count-sketch hashing on the output of GCWS hashing without hurting the accuracy much. There are many other ways to take advantage of GCWS when training deep neural networks. For example, one can apply GCWS on the outputs of the last layer to boost the accuracy of trained deep neural networks.

Weijie Zhao, Xuewu Jiao, Mingqing Hu et al.

Click-Through Rate (CTR) prediction is a crucial component in the online advertising industry. In order to produce a personalized CTR prediction, an industry-level CTR prediction model commonly takes a high-dimensional (e.g., 100 or 1000 billions of features) sparse vector (that is encoded from query keywords, user portraits, etc.) as input. As a result, the model requires Terabyte scale parameters to embed the high-dimensional input. Hierarchical distributed GPU parameter server has been proposed to enable GPU with limited memory to train the massive network by leveraging CPU main memory and SSDs as secondary storage. We identify two major challenges in the existing GPU training framework for massive-scale ad models and propose a collection of optimizations to tackle these challenges: (a) the GPU, CPU, SSD rapidly communicate with each other during the training. The connections between GPUs and CPUs are non-uniform due to the hardware topology. The data communication route should be optimized according to the hardware topology; (b) GPUs in different computing nodes frequently communicates to synchronize parameters. We are required to optimize the communications so that the distributed system can become scalable. In this paper, we propose a hardware-aware training workflow that couples the hardware topology into the algorithm design. To reduce the extensive communication between computing nodes, we introduce a $k$-step model merging algorithm for the popular Adam optimizer and provide its convergence rate in non-convex optimization. To the best of our knowledge, this is the first application of $k$-step adaptive optimization method in industrial-level CTR model training. The numerical results on real-world data confirm that the optimized system design considerably reduces the training time of the massive model, with essentially no loss in accuracy.

Xiangyi Chen, Belhal Karimi, Weijie Zhao et al.

Adaptive gradient methods including Adam, AdaGrad, and their variants have been very successful for training deep learning models, such as neural networks. Meanwhile, given the need for distributed computing, distributed optimization algorithms are rapidly becoming a focal point. With the growth of computing power and the need for using machine learning models on mobile devices, the communication cost of distributed training algorithms needs careful consideration. In this paper, we introduce novel convergent decentralized adaptive gradient methods and rigorously incorporate adaptive gradient methods into decentralized training procedures. Specifically, we propose a general algorithmic framework that can convert existing adaptive gradient methods to their decentralized counterparts. In addition, we thoroughly analyze the convergence behavior of the proposed algorithmic framework and show that if a given adaptive gradient method converges, under some specific conditions, then its decentralized counterpart is also convergent. We illustrate the benefit of our generic decentralized framework on a prototype method, i.e., AMSGrad, both theoretically and numerically.

Yan Hong, Li Niu, Jianfu Zhang et al.

In order to generate images for a given category, existing deep generative models generally rely on abundant training images. However, extensive data acquisition is expensive and fast learning ability from limited data is necessarily required in real-world applications. Also, these existing methods are not well-suited for fast adaptation to a new category. Few-shot image generation, aiming to generate images from only a few images for a new category, has attracted some research interest. In this paper, we propose a Fusing-and-Filling Generative Adversarial Network (F2GAN) to generate realistic and diverse images for a new category with only a few images. In our F2GAN, a fusion generator is designed to fuse the high-level features of conditional images with random interpolation coefficients, and then fills in attended low-level details with non-local attention module to produce a new image. Moreover, our discriminator can ensure the diversity of generated images by a mode seeking loss and an interpolation regression loss. Extensive experiments on five datasets demonstrate the effectiveness of our proposed method for few-shot image generation.

Weijie Zhao, Deping Xie, Ronglai Jia et al.

Neural networks of ads systems usually take input from multiple resources, e.g., query-ad relevance, ad features and user portraits. These inputs are encoded into one-hot or multi-hot binary features, with typically only a tiny fraction of nonzero feature values per example. Deep learning models in online advertising industries can have terabyte-scale parameters that do not fit in the GPU memory nor the CPU main memory on a computing node. For example, a sponsored online advertising system can contain more than $10^{11}$ sparse features, making the neural network a massive model with around 10 TB parameters. In this paper, we introduce a distributed GPU hierarchical parameter server for massive scale deep learning ads systems. We propose a hierarchical workflow that utilizes GPU High-Bandwidth Memory, CPU main memory and SSD as 3-layer hierarchical storage. All the neural network training computations are contained in GPUs. Extensive experiments on real-world data confirm the effectiveness and the scalability of the proposed system. A 4-node hierarchical GPU parameter server can train a model more than 2X faster than a 150-node in-memory distributed parameter server in an MPI cluster. In addition, the price-performance ratio of our proposed system is 4-9 times better than an MPI-cluster solution.

Yi Tu, Li Niu, Weijie Zhao et al.

Aesthetic image cropping is a practical but challenging task which aims at finding the best crops with the highest aesthetic quality in an image. Recently, many deep learning methods have been proposed to address this problem, but they did not reveal the intrinsic mechanism of aesthetic evaluation. In this paper, we propose an interpretable image cropping model to unveil the mystery. For each image, we use a fully convolutional network to produce an aesthetic score map, which is shared among all candidate crops during crop-level aesthetic evaluation. Then, we require the aesthetic score map to be both composition-aware and saliency-aware. In particular, the same region is assigned with different aesthetic scores based on its relative positions in different crops. Moreover, a visually salient region is supposed to have more sensitive aesthetic scores so that our network can learn to place salient objects at more proper positions. Such an aesthetic score map can be used to localize aesthetically important regions in an image, which sheds light on the composition rules learned by our model. We show the competitive performance of our model in the image cropping task on several benchmark datasets, and also demonstrate its generality in real-world applications.

Yaoyi Li, Jianfu Zhang, Weijie Zhao et al.

Recently, significant progress has been achieved in deep image matting. Most of the classical image matting methods are time-consuming and require an ideal trimap which is difficult to attain in practice. A high efficient image matting method based on a weakly annotated mask is in demand for mobile applications. In this paper, we propose a novel method based on Deep Learning and Guided Filter, called Inductive Guided Filter, which can tackle the real-time general image matting task on mobile devices. We design a lightweight hourglass network to parameterize the original Guided Filter method that takes an image and a weakly annotated mask as input. Further, the use of Gabor loss is proposed for training networks for complicated textures in image matting. Moreover, we create an image matting dataset MAT-2793 with a variety of foreground objects. Experimental results demonstrate that our proposed method massively reduces running time with robust accuracy.