Jianyu Wang

Mingkai Deng, Jianyu Wang, Cheng-Ping Hsieh et al.

Prompting has shown impressive success in enabling large pretrained language models (LMs) to perform diverse NLP tasks, especially when only few downstream data are available. Automatically finding the optimal prompt for each task, however, is challenging. Most existing work resorts to tuning soft prompt (e.g., embeddings) which falls short of interpretability, reusability across LMs, and applicability when gradients are not accessible. Discrete prompt, on the other hand, is difficult to optimize, and is often created by "enumeration (e.g., paraphrasing)-then-selection" heuristics that do not explore the prompt space systematically. This paper proposes RLPrompt, an efficient discrete prompt optimization approach with reinforcement learning (RL). RLPrompt formulates a parameter-efficient policy network that generates the desired discrete prompt after training with reward. To overcome the complexity and stochasticity of reward signals by the large LM environment, we incorporate effective reward stabilization that substantially enhances the training efficiency. RLPrompt is flexibly applicable to different types of LMs, such as masked (e.g., BERT) and left-to-right models (e.g., GPTs), for both classification and generation tasks. Experiments on few-shot classification and unsupervised text style transfer show superior performance over a wide range of existing finetuning or prompting methods. Interestingly, the resulting optimized prompts are often ungrammatical gibberish text; and surprisingly, those gibberish prompts are transferrable between different LMs to retain significant performance, indicating LM prompting may not follow human language patterns.

John Nguyen, Jianyu Wang, Kshitiz Malik et al.

An oft-cited challenge of federated learning is the presence of heterogeneity. \emph{Data heterogeneity} refers to the fact that data from different clients may follow very different distributions. \emph{System heterogeneity} refers to the fact that client devices have different system capabilities. A considerable number of federated optimization methods address this challenge. In the literature, empirical evaluations usually start federated training from random initialization. However, in many practical applications of federated learning, the server has access to proxy data for the training task that can be used to pre-train a model before starting federated training. We empirically study the impact of starting from a pre-trained model in federated learning using four standard federated learning benchmark datasets. Unsurprisingly, starting from a pre-trained model reduces the training time required to reach a target error rate and enables the training of more accurate models (up to 40\%) than is possible when starting from random initialization. Surprisingly, we also find that starting federated learning from a pre-trained initialization reduces the effect of both data and system heterogeneity. We recommend that future work proposing and evaluating federated optimization methods evaluate the performance when starting from random and pre-trained initializations. We also believe this study raises several questions for further work on understanding the role of heterogeneity in federated optimization.

Jianyu Wang, Linruize Tang, Jie Chen et al.

Nonnegative Tucker Factorization (NTF) minimizes the euclidean distance or Kullback-Leibler divergence between the original data and its low-rank approximation which often suffers from grossly corruptions or outliers and the neglect of manifold structures of data. In particular, NTF suffers from rotational ambiguity, whose solutions with and without rotation transformations are equally in the sense of yielding the maximum likelihood. In this paper, we propose three Robust Manifold NTF algorithms to handle outliers by incorporating structural knowledge about the outliers. They first applies a half-quadratic optimization algorithm to transform the problem into a general weighted NTF where the weights are influenced by the outliers. Then, we introduce the correntropy induced metric, Huber function and Cauchy function for weights respectively, to handle the outliers. Finally, we introduce a manifold regularization to overcome the rotational ambiguity of NTF. We have compared the proposed method with a number of representative references covering major branches of NTF on a variety of real-world image databases. Experimental results illustrate the effectiveness of the proposed method under two evaluation metrics (accuracy and nmi).

Xintong Liu, Jianyu Wang, Leping Xiao et al.

The non-line-of-sight imaging technique aims to reconstruct targets from multiply reflected light. For most existing methods, dense points on the relay surface are raster scanned to obtain high-quality reconstructions, which requires a long acquisition time. In this work, we propose a signal-surface collaborative regularization (SSCR) framework that provides noise-robust reconstructions with a minimal number of measurements. Using Bayesian inference, we design joint regularizations of the estimated signal, the 3D voxel-based representation of the objects, and the 2D surface-based description of the targets. To our best knowledge, this is the first work that combines regularizations in mixed dimensions for hidden targets. Experiments on synthetic and experimental datasets illustrated the efficiency and robustness of the proposed method under both confocal and non-confocal settings. We report the reconstruction of the hidden targets with complex geometric structures with only $5 \times 5$ confocal measurements from public datasets, indicating an acceleration of the conventional measurement process by a factor of 10000. Besides, the proposed method enjoys low time and memory complexities with sparse measurements. Our approach has great potential in real-time non-line-of-sight imaging applications such as rescue operations and autonomous driving.

Tianchen Deng, Guole Shen, Tong Qin et al.

Neural implicit scene representations have recently shown encouraging results in dense visual SLAM. However, existing methods produce low-quality scene reconstruction and low-accuracy localization performance when scaling up to large indoor scenes and long sequences. These limitations are mainly due to their single, global radiance field with finite capacity, which does not adapt to large scenarios. Their end-to-end pose networks are also not robust enough with the growth of cumulative errors in large scenes. To this end, we introduce PLGSLAM, a neural visual SLAM system capable of high-fidelity surface reconstruction and robust camera tracking in real-time. To handle large-scale indoor scenes, PLGSLAM proposes a progressive scene representation method which dynamically allocates new local scene representation trained with frames within a local sliding window. This allows us to scale up to larger indoor scenes and improves robustness (even under pose drifts). In local scene representation, PLGSLAM utilizes tri-planes for local high-frequency features with multi-layer perceptron (MLP) networks for the low-frequency feature, achieving smoothness and scene completion in unobserved areas. Moreover, we propose local-to-global bundle adjustment method with a global keyframe database to address the increased pose drifts on long sequences. Experimental results demonstrate that PLGSLAM achieves state-of-the-art scene reconstruction results and tracking performance across various datasets and scenarios (both in small and large-scale indoor environments). The code is open-sourced at https://github.com/dtc111111/plgslam.

Kaiyan Zhang, Jianyu Wang, Ermo Hua et al.

With the advancement of language models (LMs), their exposure to private data is increasingly inevitable, and their deployment (especially for smaller ones) on personal devices, such as PCs and smartphones, has become a prevailing trend. In contexts laden with user information, enabling models to both safeguard user privacy and execute commands efficiently emerges as an essential research imperative. In this paper, we propose CoGenesis, a collaborative generation framework integrating large (hosted on cloud infrastructure) and small models (deployed on local devices) to address privacy concerns logically. Initially, we design a pipeline to create personalized writing instruction datasets enriched with extensive context details as the testbed of this research issue. Subsequently, we introduce two variants of CoGenesis based on sketch and logits respectively. Our experimental findings, based on our synthesized dataset and two additional open-source datasets, indicate that: 1) Large-scale models perform well when provided with user context but struggle in the absence of such context. 2) While specialized smaller models fine-tuned on the synthetic dataset show promise, they still lag behind their larger counterparts. 3) Our CoGenesis framework, utilizing mixed-scale models, showcases competitive performance, providing a feasible solution to privacy issues.

Ziyao Wang, Jianyu Wang, Ang Li

The theoretical landscape of federated learning (FL) undergoes rapid evolution, but its practical application encounters a series of intricate challenges, and hyperparameter optimization is one of these critical challenges. Amongst the diverse adjustments in hyperparameters, the adaptation of the learning rate emerges as a crucial component, holding the promise of significantly enhancing the efficacy of FL systems. In response to this critical need, this paper presents FedHyper, a novel hypergradient-based learning rate adaptation algorithm specifically designed for FL. FedHyper serves as a universal learning rate scheduler that can adapt both global and local rates as the training progresses. In addition, FedHyper not only showcases unparalleled robustness to a spectrum of initial learning rate configurations but also significantly alleviates the necessity for laborious empirical learning rate adjustments. We provide a comprehensive theoretical analysis of FedHyper's convergence rate and conduct extensive experiments on vision and language benchmark datasets. The results demonstrate that FEDHYPER consistently converges 1.1-3x faster than FedAvg and the competing baselines while achieving superior final accuracy. Moreover, FedHyper catalyzes a remarkable surge in accuracy, augmenting it by up to 15% compared to FedAvg under suboptimal initial learning rate settings.

Kaiyan Zhang, Jianyu Wang, Ning Ding et al.

Large Language Models (LLMs) exhibit impressive capabilities across various applications but encounter substantial challenges such as high inference latency, considerable training costs, and the generation of hallucinations. Collaborative decoding between large and small language models (SLMs) presents a promising strategy to mitigate these issues through methods including speculative decoding, contrastive decoding, and emulator or proxy fine-tuning. However, the specifics of such collaborations, particularly from a unified perspective, remain largely unexplored. Inspired by dual-process cognitive theory, we propose a unified framework in this paper, termed Fast and Slow Generating (FS-GEN). Within this framework, LLMs (sometimes along with SLMs) are categorized as System 2 (slow and deliberate), while independent SLMs are designated as System 1 (fast and intuitive). We provide a comprehensive analysis of these collaborative methodologies, elucidating their common properties and shedding light on the differential knowledge capabilities of System 2 versus System 1 through the FS-GEN framework. Our findings indicate that only a small proportion of collaborative interactions (approximately less than 20\% in most instances) are necessary across various methods. These interactions between System 1 and System 2 conform to a scaling law related to the parameter ratios, enabling predictable collaboration. Furthermore, we explore the specific conditions under which collaboration proves most effective, particularly from an uncertainty perspective, offering novel insights that may guide future optimization efforts. Our research underscores that the fundamental distinction between System 1 and System 2 lies in the uncertainty of next token predictions, where interventions by System 2 are crucial to support System 1. Code for Reproduction: https://github.com/TsinghuaC3I/FS-GEN

Rui Ye, Mingkai Xu, Jianyu Wang et al.

This work considers the category distribution heterogeneity in federated learning. This issue is due to biased labeling preferences at multiple clients and is a typical setting of data heterogeneity. To alleviate this issue, most previous works consider either regularizing local models or fine-tuning the global model, while they ignore the adjustment of aggregation weights and simply assign weights based on the dataset size. However, based on our empirical observations and theoretical analysis, we find that the dataset size is not optimal and the discrepancy between local and global category distributions could be a beneficial and complementary indicator for determining aggregation weights. We thus propose a novel aggregation method, Federated Learning with Discrepancy-aware Collaboration (FedDisco), whose aggregation weights not only involve both the dataset size and the discrepancy value, but also contribute to a tighter theoretical upper bound of the optimization error. FedDisco also promotes privacy-preservation, communication and computation efficiency, as well as modularity. Extensive experiments show that our FedDisco outperforms several state-of-the-art methods and can be easily incorporated with many existing methods to further enhance the performance. Our code will be available at https://github.com/MediaBrain-SJTU/FedDisco.

Yantao Lu, Yunhan Jia, Jianyu Wang et al.

Neural networks are known to be vulnerable to carefully crafted adversarial examples, and these malicious samples often transfer, i.e., they remain adversarial even against other models. Although great efforts have been delved into the transferability across models, surprisingly, less attention has been paid to the cross-task transferability, which represents the real-world cybercriminal's situation, where an ensemble of different defense/detection mechanisms need to be evaded all at once. In this paper, we investigate the transferability of adversarial examples across a wide range of real-world computer vision tasks, including image classification, object detection, semantic segmentation, explicit content detection, and text detection. Our proposed attack minimizes the ``dispersion'' of the internal feature map, which overcomes existing attacks' limitation of requiring task-specific loss functions and/or probing a target model. We conduct evaluation on open source detection and segmentation models as well as four different computer vision tasks provided by Google Cloud Vision (GCV) APIs, to show how our approach outperforms existing attacks by degrading performance of multiple CV tasks by a large margin with only modest perturbations linf=16.

Cihang Xie, Zhishuai Zhang, Yuyin Zhou et al.

Though CNNs have achieved the state-of-the-art performance on various vision tasks, they are vulnerable to adversarial examples --- crafted by adding human-imperceptible perturbations to clean images. However, most of the existing adversarial attacks only achieve relatively low success rates under the challenging black-box setting, where the attackers have no knowledge of the model structure and parameters. To this end, we propose to improve the transferability of adversarial examples by creating diverse input patterns. Instead of only using the original images to generate adversarial examples, our method applies random transformations to the input images at each iteration. Extensive experiments on ImageNet show that the proposed attack method can generate adversarial examples that transfer much better to different networks than existing baselines. By evaluating our method against top defense solutions and official baselines from NIPS 2017 adversarial competition, the enhanced attack reaches an average success rate of 73.0%, which outperforms the top-1 attack submission in the NIPS competition by a large margin of 6.6%. We hope that our proposed attack strategy can serve as a strong benchmark baseline for evaluating the robustness of networks to adversaries and the effectiveness of different defense methods in the future. Code is available at https://github.com/cihangxie/DI-2-FGSM.

Cihang Xie, Jianyu Wang, Zhishuai Zhang et al.

Convolutional neural networks have demonstrated high accuracy on various tasks in recent years. However, they are extremely vulnerable to adversarial examples. For example, imperceptible perturbations added to clean images can cause convolutional neural networks to fail. In this paper, we propose to utilize randomization at inference time to mitigate adversarial effects. Specifically, we use two randomization operations: random resizing, which resizes the input images to a random size, and random padding, which pads zeros around the input images in a random manner. Extensive experiments demonstrate that the proposed randomization method is very effective at defending against both single-step and iterative attacks. Our method provides the following advantages: 1) no additional training or fine-tuning, 2) very few additional computations, 3) compatible with other adversarial defense methods. By combining the proposed randomization method with an adversarially trained model, it achieves a normalized score of 0.924 (ranked No.2 among 107 defense teams) in the NIPS 2017 adversarial examples defense challenge, which is far better than using adversarial training alone with a normalized score of 0.773 (ranked No.56). The code is public available at https://github.com/cihangxie/NIPS2017_adv_challenge_defense.

Bairu Hou, Qibin Chen, Jianyu Wang et al.

With the rapid scaling of large language models (LLMs), structured pruning has become a widely used technique to learn efficient, smaller models from larger ones, delivering superior performance compared to training similarly sized models from scratch. In this paper, we move beyond the traditional static pruning approach of determining a fixed pruning mask for a model, and propose a dynamic approach to structured pruning. In our method, the pruning mask is input-dependent and adapts dynamically based on the information described in a user instruction. Our approach, termed "instruction-following pruning", introduces a sparse mask predictor that takes the user instruction as input and dynamically selects the most relevant model parameters for the given task. To identify and activate effective parameters, we jointly optimize the sparse mask predictor and the LLM, leveraging both instruction-following data and the pre-training corpus. Experimental results demonstrate the effectiveness of our approach on a wide range of evaluation benchmarks. For example, our 3B activated model improves over the 3B dense model by 5-8 points of absolute margin on domains such as math and coding, and rivals the performance of a 9B model.

Ruifeng Yuan, Chenghao Xiao, Sicong Leng et al.

Reinforcement learning has proven its effectiveness in enhancing the reasoning capabilities of large language models. Recent research efforts have progressively extended this paradigm to multimodal reasoning tasks. Due to the inherent complexity and diversity of multimodal tasks, especially in semantic content and problem formulations, existing models often exhibit unstable performance across various domains and difficulty levels. To address these limitations, we propose VL-Cogito, an advanced multimodal reasoning model trained via a novel multi-stage Progressive Curriculum Reinforcement Learning (PCuRL) framework. PCuRL systematically guides the model through tasks of gradually increasing difficulty, substantially improving its reasoning abilities across diverse multimodal contexts. The framework introduces two key innovations: (1) an online difficulty soft weighting mechanism, dynamically adjusting training difficulty across successive RL training stages; and (2) a dynamic length reward mechanism, which encourages the model to adaptively regulate its reasoning path length according to task complexity, thus balancing reasoning efficiency with correctness. Experimental evaluations demonstrate that VL-Cogito consistently matches or surpasses existing reasoning-oriented models across mainstream multimodal benchmarks spanning mathematics, science, logic, and general understanding, validating the effectiveness of our approach.

Mark Lee, Tom Gunter, Chang Lan et al.

We design and implement AXLearn, a production deep learning system that facilitates scalable and high-performance training of large deep learning models. Compared to other state-of-the-art deep learning systems, AXLearn has a unique focus on modularity and support for heterogeneous hardware infrastructure. AXLearn's internal interfaces between software components follow strict encapsulation, allowing different components to be assembled to facilitate rapid model development and experimentation on heterogeneous compute infrastructure. We introduce a novel method of quantifying modularity via Lines-of-Code (LoC)-complexity, which demonstrates how our system maintains constant complexity as we scale the components in the system, compared to linear or quadratic complexity in other systems. This allows integrating features such as Rotary Position Embeddings (RoPE) into AXLearn across hundred of modules with just 10 lines of code, compared to hundreds as required in other systems. At the same time, AXLearn maintains equivalent performance compared to state-of-the-art training systems. Finally, we share our experience in the development and operation of AXLearn.

Tao Yu, Congzheng Song, Jianyu Wang et al.

Asynchronous protocols have been shown to improve the scalability of federated learning (FL) with a massive number of clients. Meanwhile, momentum-based methods can achieve the best model quality in synchronous FL. However, naively applying momentum in asynchronous FL algorithms leads to slower convergence and degraded model performance. It is still unclear how to effective combinie these two techniques together to achieve a win-win. In this paper, we find that asynchrony introduces implicit bias to momentum updates. In order to address this problem, we propose momentum approximation that minimizes the bias by finding an optimal weighted average of all historical model updates. Momentum approximation is compatible with secure aggregation as well as differential privacy, and can be easily integrated in production FL systems with a minor communication and storage cost. We empirically demonstrate that on benchmark FL datasets, momentum approximation can achieve $1.15 \textrm{--}4\times$ speed up in convergence compared to naively combining asynchronous FL with momentum.

Yijun Wei, Jianyu Wang, Leping Xiao et al.

Non-line-of-sight (NLOS) imaging seeks to reconstruct hidden objects by analyzing reflections from intermediary surfaces. Existing methods typically model both the measurement data and the hidden scene in three dimensions, overlooking the inherently two-dimensional nature of most hidden objects. This oversight leads to high computational costs and substantial memory consumption, limiting practical applications and making real-time, high-resolution NLOS imaging on lightweight devices challenging. In this paper, we introduce a novel approach that represents the hidden scene using two-dimensional functions and employs a Quasi-Fresnel transform to establish a direct inversion formula between the measurement data and the hidden scene. This transformation leverages the two-dimensional characteristics of the problem to significantly reduce computational complexity and memory requirements. Our algorithm efficiently performs fast transformations between these two-dimensional aggregated data, enabling rapid reconstruction of hidden objects with minimal memory usage. Compared to existing methods, our approach reduces runtime and memory demands by several orders of magnitude while maintaining imaging quality. The substantial reduction in memory usage not only enhances computational efficiency but also enables NLOS imaging on lightweight devices such as mobile and embedded systems. We anticipate that this method will facilitate real-time, high-resolution NLOS imaging and broaden its applicability across a wider range of platforms.

Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier et al.

In this work, we discuss building performant Multimodal Large Language Models (MLLMs). In particular, we study the importance of various architecture components and data choices. Through careful and comprehensive ablations of the image encoder, the vision language connector, and various pre-training data choices, we identified several crucial design lessons. For example, we demonstrate that for large-scale multimodal pre-training using a careful mix of image-caption, interleaved image-text, and text-only data is crucial for achieving state-of-the-art (SOTA) few-shot results across multiple benchmarks, compared to other published pre-training results. Further, we show that the image encoder together with image resolution and the image token count has substantial impact, while the vision-language connector design is of comparatively negligible importance. By scaling up the presented recipe, we build MM1, a family of multimodal models up to 30B parameters, including both dense models and mixture-of-experts (MoE) variants, that are SOTA in pre-training metrics and achieve competitive performance after supervised fine-tuning on a range of established multimodal benchmarks. Thanks to large-scale pre-training, MM1 enjoys appealing properties such as enhanced in-context learning, and multi-image reasoning, enabling few-shot chain-of-thought prompting.

JianYu Wang, Xiao-Lei Zhang

Nonnegative matrix factorization (NMF) based topic modeling methods do not rely on model- or data-assumptions much. However, they are usually formulated as difficult optimization problems, which may suffer from bad local minima and high computational complexity. In this paper, we propose a deep NMF (DNMF) topic modeling framework to alleviate the aforementioned problems. It first applies an unsupervised deep learning method to learn latent hierarchical structures of documents, under the assumption that if we could learn a good representation of documents by, e.g. a deep model, then the topic word discovery problem can be boosted. Then, it takes the output of the deep model to constrain a topic-document distribution for the discovery of the discriminant topic words, which not only improves the efficacy but also reduces the computational complexity over conventional unsupervised NMF methods. We constrain the topic-document distribution in three ways, which takes the advantages of the three major sub-categories of NMF -- basic NMF, structured NMF, and constrained NMF respectively. To overcome the weaknesses of deep neural networks in unsupervised topic modeling, we adopt a non-neural-network deep model -- multilayer bootstrap network. To our knowledge, this is the first time that a deep NMF model is used for unsupervised topic modeling. We have compared the proposed method with a number of representative references covering major branches of topic modeling on a variety of real-world text corpora. Experimental results illustrate the effectiveness of the proposed method under various evaluation metrics.

Jianyu Wang, Shanzheng Guan, Shupei Liu et al.

Multichannel blind audio source separation aims to recover the latent sources from their multichannel mixtures without supervised information. One state-of-the-art blind audio source separation method, named independent low-rank matrix analysis (ILRMA), unifies independent vector analysis (IVA) and nonnegative matrix factorization (NMF). However, the spectra matrix produced from NMF may not find a compact spectral basis. It may not guarantee the identifiability of each source as well. To address this problem, here we propose to enhance the identifiability of the source model by a minimum-volume prior distribution. We further regularize a multichannel NMF (MNMF) and ILRMA respectively with the minimum-volume regularizer. The proposed methods maximize the posterior distribution of the separated sources, which ensures the stability of the convergence. Experimental results demonstrate the effectiveness of the proposed methods compared with auxiliary independent vector analysis, MNMF, ILRMA and its extensions.

Yuanpeng Li, Liang Zhao, Jianyu Wang et al.

Compositional generalization is a basic mechanism in human language learning, but current neural networks lack such ability. In this paper, we conduct fundamental research for encoding compositionality in neural networks. Conventional methods use a single representation for the input sentence, making it hard to apply prior knowledge of compositionality. In contrast, our approach leverages such knowledge with two representations, one generating attention maps, and the other mapping attended input words to output symbols. We reduce the entropy in each representation to improve generalization. Our experiments demonstrate significant improvements over the conventional methods in five NLP tasks including instruction learning and machine translation. In the SCAN domain, it boosts accuracies from 14.0% to 98.8% in Jump task, and from 92.0% to 99.7% in TurnLeft task. It also beats human performance on a few-shot learning task. We hope the proposed approach can help ease future research towards human-level compositional language learning.

Haichao Zhang, Jianyu Wang

We introduce a feature scattering-based adversarial training approach for improving model robustness against adversarial attacks. Conventional adversarial training approaches leverage a supervised scheme (either targeted or non-targeted) in generating attacks for training, which typically suffer from issues such as label leaking as noted in recent works. Differently, the proposed approach generates adversarial images for training through feature scattering in the latent space, which is unsupervised in nature and avoids label leaking. More importantly, this new approach generates perturbed images in a collaborative fashion, taking the inter-sample relationships into consideration. We conduct analysis on model robustness and demonstrate the effectiveness of the proposed approach through extensively experiments on different datasets compared with state-of-the-art approaches.

Haichao Zhang, Jianyu Wang

Conventional adversarial training methods using attacks that manipulate the pixel value directly and individually, leading to models that are less robust in face of spatial transformation-based attacks. In this paper, we propose a joint adversarial training method that incorporates both spatial transformation-based and pixel-value based attacks for improving model robustness. We introduce a spatial transformation-based attack with an explicit notion of budget and develop an algorithm for spatial attack generation. We further integrate both pixel and spatial attacks into one generation model and show how to leverage the complementary strengths of each other in training for improving the overall model robustness. Extensive experimental results on different benchmark datasets compared with state-of-the-art methods verified the effectiveness of the proposed method.

Haichao Zhang, Jianyu Wang

Object detection is an important vision task and has emerged as an indispensable component in many vision system, rendering its robustness as an increasingly important performance factor for practical applications. While object detection models have been demonstrated to be vulnerable against adversarial attacks by many recent works, very few efforts have been devoted to improving their robustness. In this work, we take an initial attempt towards this direction. We first revisit and systematically analyze object detectors and many recently developed attacks from the perspective of model robustness. We then present a multi-task learning perspective of object detection and identify an asymmetric role of task losses. We further develop an adversarial training approach which can leverage the multiple sources of attacks for improving the robustness of detection models. Extensive experiments on PASCAL-VOC and MS-COCO verified the effectiveness of the proposed approach.

Yuanpeng Li, Yi Yang, Jianyu Wang et al.

Acquiring a large vocabulary is an important aspect of human intelligence. Onecommon approach for human to populating vocabulary is to learn words duringreading or listening, and then use them in writing or speaking. This ability totransfer from input to output is natural for human, but it is difficult for machines.Human spontaneously performs this knowledge transfer in complicated multimodaltasks, such as Visual Question Answering (VQA). In order to approach human-levelArtificial Intelligence, we hope to equip machines with such ability. Therefore, toaccelerate this research, we propose a newzero-shot transfer VQA(ZST-VQA)dataset by reorganizing the existing VQA v1.0 dataset in the way that duringtraining, some words appear only in one module (i.e. questions) but not in theother (i.e. answers). In this setting, an intelligent model should understand andlearn the concepts from one module (i.e. questions), and at test time, transfer themto the other (i.e. predict the concepts as answers). We conduct evaluation on thisnew dataset using three existing state-of-the-art VQA neural models. Experimentalresults show a significant drop in performance on this dataset, indicating existingmethods do not address the zero-shot transfer problem. Besides, our analysis findsthat this may be caused by the implicit bias learned during training.

Alexey Kurakin, Ian Goodfellow, Samy Bengio et al.

To accelerate research on adversarial examples and robustness of machine learning classifiers, Google Brain organized a NIPS 2017 competition that encouraged researchers to develop new methods to generate adversarial examples as well as to develop new ways to defend against them. In this chapter, we describe the structure and organization of the competition and the solutions developed by several of the top-placing teams.

Jianyu Wang, Zhishuai Zhang, Cihang Xie et al.

It is very attractive to formulate vision in terms of pattern theory \cite{Mumford2010pattern}, where patterns are defined hierarchically by compositions of elementary building blocks. But applying pattern theory to real world images is currently less successful than discriminative methods such as deep networks. Deep networks, however, are black-boxes which are hard to interpret and can easily be fooled by adding occluding objects. It is natural to wonder whether by better understanding deep networks we can extract building blocks which can be used to develop pattern theoretic models. This motivates us to study the internal representations of a deep network using vehicle images from the PASCAL3D+ dataset. We use clustering algorithms to study the population activities of the features and extract a set of visual concepts which we show are visually tight and correspond to semantic parts of vehicles. To analyze this we annotate these vehicles by their semantic parts to create a new dataset, VehicleSemanticParts, and evaluate visual concepts as unsupervised part detectors. We show that visual concepts perform fairly well but are outperformed by supervised discriminative methods such as Support Vector Machines (SVM). We next give a more detailed analysis of visual concepts and how they relate to semantic parts. Following this, we use the visual concepts as building blocks for a simple pattern theoretical model, which we call compositional voting. In this model several visual concepts combine to detect semantic parts. We show that this approach is significantly better than discriminative methods like SVM and deep networks trained specifically for semantic part detection. Finally, we return to studying occlusion by creating an annotated dataset with occlusion, called VehicleOcclusion, and show that compositional voting outperforms even deep networks when the amount of occlusion becomes large.

Zhishuai Zhang, Cihang Xie, Jianyu Wang et al.

In this paper, we study the task of detecting semantic parts of an object, e.g., a wheel of a car, under partial occlusion. We propose that all models should be trained without seeing occlusions while being able to transfer the learned knowledge to deal with occlusions. This setting alleviates the difficulty in collecting an exponentially large dataset to cover occlusion patterns and is more essential. In this scenario, the proposal-based deep networks, like RCNN-series, often produce unsatisfactory results, because both the proposal extraction and classification stages may be confused by the irrelevant occluders. To address this, [25] proposed a voting mechanism that combines multiple local visual cues to detect semantic parts. The semantic parts can still be detected even though some visual cues are missing due to occlusions. However, this method is manually-designed, thus is hard to be optimized in an end-to-end manner. In this paper, we present DeepVoting, which incorporates the robustness shown by [25] into a deep network, so that the whole pipeline can be jointly optimized. Specifically, it adds two layers after the intermediate features of a deep network, e.g., the pool-4 layer of VGGNet. The first layer extracts the evidence of local visual cues, and the second layer performs a voting mechanism by utilizing the spatial relationship between visual cues and semantic parts. We also propose an improved version DeepVoting+ by learning visual cues from context outside objects. In experiments, DeepVoting achieves significantly better performance than several baseline methods, including Faster-RCNN, for semantic part detection under occlusion. In addition, DeepVoting enjoys explainability as the detection results can be diagnosed via looking up the voting cues.

Jianyu Wang, Cihang Xie, Zhishuai Zhang et al.

In this paper, we address the task of detecting semantic parts on partially occluded objects. We consider a scenario where the model is trained using non-occluded images but tested on occluded images. The motivation is that there are infinite number of occlusion patterns in real world, which cannot be fully covered in the training data. So the models should be inherently robust and adaptive to occlusions instead of fitting / learning the occlusion patterns in the training data. Our approach detects semantic parts by accumulating the confidence of local visual cues. Specifically, the method uses a simple voting method, based on log-likelihood ratio tests and spatial constraints, to combine the evidence of local cues. These cues are called visual concepts, which are derived by clustering the internal states of deep networks. We evaluate our voting scheme on the VehicleSemanticPart dataset with dense part annotations. We randomly place two, three or four irrelevant objects onto the target object to generate testing images with various occlusions. Experiments show that our algorithm outperforms several competitors in semantic part detection when occlusions are present.

Cihang Xie, Jianyu Wang, Zhishuai Zhang et al.

It has been well demonstrated that adversarial examples, i.e., natural images with visually imperceptible perturbations added, generally exist for deep networks to fail on image classification. In this paper, we extend adversarial examples to semantic segmentation and object detection which are much more difficult. Our observation is that both segmentation and detection are based on classifying multiple targets on an image (e.g., the basic target is a pixel or a receptive field in segmentation, and an object proposal in detection), which inspires us to optimize a loss function over a set of pixels/proposals for generating adversarial perturbations. Based on this idea, we propose a novel algorithm named Dense Adversary Generation (DAG), which generates a large family of adversarial examples, and applies to a wide range of state-of-the-art deep networks for segmentation and detection. We also find that the adversarial perturbations can be transferred across networks with different training data, based on different architectures, and even for different recognition tasks. In particular, the transferability across networks with the same architecture is more significant than in other cases. Besides, summing up heterogeneous perturbations often leads to better transfer performance, which provides an effective method of black-box adversarial attack.

Jianyu Wang, Zhishuai Zhang, Cihang Xie et al.

We address the key question of how object part representations can be found from the internal states of CNNs that are trained for high-level tasks, such as object classification. This work provides a new unsupervised method to learn semantic parts and gives new understanding of the internal representations of CNNs. Our technique is based on the hypothesis that semantic parts are represented by populations of neurons rather than by single filters. We propose a clustering technique to extract part representations, which we call Visual Concepts. We show that visual concepts are semantically coherent in that they represent semantic parts, and visually coherent in that corresponding image patches appear very similar. Also, visual concepts provide full spatial coverage of the parts of an object, rather than a few sparse parts as is typically found in keypoint annotations. Furthermore, We treat single visual concept as part detector and evaluate it for keypoint detection using the PASCAL3D+ dataset and for part detection using our newly annotated ImageNetPart dataset. The experiments demonstrate that visual concepts can be used to detect parts. We also show that some visual concepts respond to several semantic parts, provided these parts are visually similar. Thus visual concepts have the essential properties: semantic meaning and detection capability. Note that our ImageNetPart dataset gives rich part annotations which cover the whole object, making it useful for other part-related applications.

Jianyu Wang, Alan Yuille

In this paper, we study the problem of semantic part segmentation for animals. This is more challenging than standard object detection, object segmentation and pose estimation tasks because semantic parts of animals often have similar appearance and highly varying shapes. To tackle these challenges, we build a mixture of compositional models to represent the object boundary and the boundaries of semantic parts. And we incorporate edge, appearance, and semantic part cues into the compositional model. Given part-level segmentation annotation, we develop a novel algorithm to learn a mixture of compositional models under various poses and viewpoints for certain animal classes. Furthermore, a linear complexity algorithm is offered for efficient inference of the compositional model using dynamic programming. We evaluate our method for horse and cow using a newly annotated dataset on Pascal VOC 2010 which has pixelwise part labels. Experimental results demonstrate the effectiveness of our method.