Research papers and code for "Xiang Li":
Recent progress in biomedical image segmentation based on deep convolutional neural networks (CNNs) has drawn much attention. However, its vulnerability towards adversarial samples cannot be overlooked. This paper is the first one that discovers that all the CNN-based state-of-the-art biomedical image segmentation models are sensitive to adversarial perturbations. This limits the deployment of these methods in safety-critical biomedical fields. In this paper, we discover that global spatial dependencies and global contextual information in a biomedical image can be exploited to defend against adversarial attacks. To this end, non-local context encoder (NLCE) is proposed to model short- and long range spatial dependencies and encode global contexts for strengthening feature activations by channel-wise attention. The NLCE modules enhance the robustness and accuracy of the non-local context encoding network (NLCEN), which learns robust enhanced pyramid feature representations with NLCE modules, and then integrates the information across different levels. Experiments on both lung and skin lesion segmentation datasets have demonstrated that NLCEN outperforms any other state-of-the-art biomedical image segmentation methods against adversarial attacks. In addition, NLCE modules can be applied to improve the robustness of other CNN-based biomedical image segmentation methods.

* Accepted by AAAI2019 as oral presentation
Click to Read Paper and Get Code
Explaining the prediction of deep neural networks (DNNs) and semantic image compression are two active research areas of deep learning with a numerous of applications in decision-critical systems, such as surveillance cameras, drones and self-driving cars, where interpretable decision is critical and storage/network bandwidth is limited. In this paper, we propose a novel end-to-end Neural Image Compression and Explanation (NICE) framework that learns to (1) explain the prediction of convolutional neural networks (CNNs), and (2) subsequently compress the input images for efficient storage or transmission. Specifically, NICE generates a sparse mask over an input image by attaching a stochastic binary gate to each pixel of the image, whose parameters are learned through the interaction with the CNN classifier to be explained. The generated mask is able to capture the saliency of each pixel measured by its influence to the final prediction of CNN; it can also be used to produce a mixed-resolution image, where important pixels maintain their original high resolution and insignificant background pixels are subsampled to a low resolution. The produced images achieve a high compression rate (e.g., about 0.6x of original image file size), while retaining a similar classification accuracy. Extensive experiments across multiple image classification benchmarks demonstrate the superior performance of NICE compared to the state-of-the-art methods in terms of explanation quality and image compression rate.

Click to Read Paper and Get Code
Existing methods for AI-generated artworks still struggle with generating high-quality stylized content, where high-level semantics are preserved, or separating fine-grained styles from various artists. We propose a novel Generative Adversarial Disentanglement Network which can fully decompose complex anime illustrations into style and content. Training such model is challenging, since given a style, various content data may exist but not the other way round. In particular, we disentangle two complementary factors of variations, where one of the factors is labelled. Our approach is divided into two stages, one that encodes an input image into a style independent content, and one based on a dual-conditional generator. We demonstrate the ability to generate high-fidelity anime portraits with a fixed content and a large variety of styles from over a thousand artists, and vice versa, using a single end-to-end network and with applications in style transfer. We show this unique capability as well as superior output to the current state-of-the-art.

* Submitted to NeurIPS 2019
Click to Read Paper and Get Code
Deep neural networks (DNNs) have been enormously successful across a variety of prediction tasks. However, recent research shows that DNNs are particularly vulnerable to adversarial attacks, which poses a serous threat to their applications in security-sensitive systems. In this paper, we propose a simple yet effective defense algorithm Defense-VAE that uses variational autoencoder (VAE) to purge adversarial perturbations from contaminated images. The proposed method is generic and can defend white-box and black-box attacks without the need of retraining the original CNN classifiers, and can further strengthen the defense by retraining CNN or end-to-end finetuning the whole pipeline. In addition, the proposed method is very efficient compared to the optimization-based alternatives, such as Defense-GAN, since no iterative optimization is needed for online prediction. Extensive experiments on MNIST, Fashion-MNIST, CelebA and CIFAR-10 demonstrate the superior defense accuracy of Defense-VAE compared to Defense-GAN, while being 50x faster than the latter. This makes Defense-VAE widely deployable in real-time security-sensitive systems. We plan to open source our implementation to facilitate the research in this area.

Click to Read Paper and Get Code
Deep learning-based style transfer between images has recently become a popular area of research. A common way of encoding "style" is through a feature representation based on the Gram matrix of features extracted by some pre-trained neural network or some other form of feature statistics. Such a definition is based on an arbitrary human decision and may not best capture what a style really is. In trying to gain a better understanding of "style", we propose a metric learning-based method to explicitly encode the style of an artwork. In particular, our definition of style captures the differences between artists, as shown by classification performances, and such that the style representation can be interpreted, manipulated and visualized through style-conditioned image generation through a Generative Adversarial Network. We employ this method to explore the style space of anime portrait illustrations.

Click to Read Paper and Get Code
Generative adversarial networks (GANs) are highly effective unsupervised learning frameworks that can generate very sharp data, even for data such as images with complex, highly multimodal distributions. However GANs are known to be very hard to train, suffering from problems such as mode collapse and disturbing visual artifacts. Batch normalization (BN) techniques have been introduced to address the training. Though BN accelerates the training in the beginning, our experiments show that the use of BN can be unstable and negatively impact the quality of the trained model. The evaluation of BN and numerous other recent schemes for improving GAN training is hindered by the lack of an effective objective quality measure for GAN models. To address these issues, we first introduce a weight normalization (WN) approach for GAN training that significantly improves the stability, efficiency and the quality of the generated samples. To allow a methodical evaluation, we introduce squared Euclidean reconstruction error on a test set as a new objective measure, to assess training performance in terms of speed, stability, and quality of generated samples. Our experiments with a standard DCGAN architecture on commonly used datasets (CelebA, LSUN bedroom, and CIFAR-10) indicate that training using WN is generally superior to BN for GANs, achieving 10% lower mean squared loss for reconstruction and significantly better qualitative results than BN. We further demonstrate the stability of WN on a 21-layer ResNet trained with the CelebA data set. The code for this paper is available at https://github.com/stormraiser/gan-weightnorm-resnet

* v3 rejected by NIPS 2017, updated and re-submitted to CVPR 2018. v4: add experiments with ResNet and like to new code
Click to Read Paper and Get Code
Hyperspectral image (HSI) classification, which aims to assign an accurate label for hyperspectral pixels, has drawn great interest in recent years. Although low rank representation (LRR) has been used to classify HSI, its ability to segment each class from the whole HSI data has not been exploited fully yet. LRR has a good capacity to capture the underlying lowdimensional subspaces embedded in original data. However, there are still two drawbacks for LRR. First, LRR does not consider the local geometric structure within data, which makes the local correlation among neighboring data easily ignored. Second, the representation obtained by solving LRR is not discriminative enough to separate different data. In this paper, a novel locality and structure regularized low rank representation (LSLRR) model is proposed for HSI classification. To overcome the above limitations, we present locality constraint criterion (LCC) and structure preserving strategy (SPS) to improve the classical LRR. Specifically, we introduce a new distance metric, which combines both spatial and spectral features, to explore the local similarity of pixels. Thus, the global and local structures of HSI data can be exploited sufficiently. Besides, we propose a structure constraint to make the representation have a near block-diagonal structure. This helps to determine the final classification labels directly. Extensive experiments have been conducted on three popular HSI datasets. And the experimental results demonstrate that the proposed LSLRR outperforms other state-of-the-art methods.

* IEEE Transactions on Geoscience and Remote Sensing ( Volume: 57 , Issue: 2 , Feb. 2019 )
* 14 pages, 7 figures, TGRS2019
Click to Read Paper and Get Code
The Convolutional Neural Networks (CNNs) generate the feature representation of complex objects by collecting hierarchical and different parts of semantic sub-features. These sub-features can usually be distributed in grouped form in the feature vector of each layer, representing various semantic entities. However, the activation of these sub-features is often spatially affected by similar patterns and noisy backgrounds, resulting in erroneous localization and identification. We propose a Spatial Group-wise Enhance (SGE) module that can adjust the importance of each sub-feature by generating an attention factor for each spatial location in each semantic group, so that every individual group can autonomously enhance its learnt expression and suppress possible noise. The attention factors are only guided by the similarities between the global and local feature descriptors inside each group, thus the design of SGE module is extremely lightweight with \emph{almost no extra parameters and calculations}. Despite being trained with only category supervisions, the SGE component is extremely effective in highlighting multiple active areas with various high-order semantics (such as the dog's eyes, nose, etc.). When integrated with popular CNN backbones, SGE can significantly boost the performance of image recognition tasks. Specifically, based on ResNet50 backbones, SGE achieves 1.2\% Top-1 accuracy improvement on the ImageNet benchmark and 1.0$\sim$2.0\% AP gain on the COCO benchmark across a wide range of detectors (Faster/Mask/Cascade RCNN and RetinaNet). Codes and pretrained models are available at https://github.com/implus/PytorchInsight.

* Code available at: https://github.com/implus/PytorchInsight
Click to Read Paper and Get Code
We propose and study a general framework for regularized Markov decision processes (MDPs) where the goal is to find an optimal policy that maximizes the expected discounted total reward plus a policy regularization term. The extant entropy-regularized MDPs can be cast into our framework. Moreover, under our framework there are many regularization terms can bring multi-modality and sparsity which are potentially useful in reinforcement learning. In particular, we present sufficient and necessary conditions that induce a sparse optimal policy. We also conduct a full mathematical analysis of the proposed regularized MDPs, including the optimality condition, performance error and sparseness control. We provide a generic method to devise regularization forms and propose off-policy actor-critic algorithms in complex environment settings. We empirically analyze the statistical properties of optimal policies and compare the performance of different sparse regularization forms in discrete and continuous environments.

Click to Read Paper and Get Code
Subsampled Newton methods approximate Hessian matrices through subsampling techniques, alleviating the cost of forming Hessian matrices but using sufficient curvature information. However, previous results require $\Omega (d)$ samples to approximate Hessians, where $d$ is the dimension of data points, making it less practically feasible for high-dimensional data. The situation is deteriorated when $d$ is comparably as large as the number of data points $n$, which requires to take the whole dataset into account, making subsampling useless. This paper theoretically justifies the effectiveness of subsampled Newton methods on high dimensional data. Specifically, we prove only $\widetilde{\Theta}(d^\gamma_{\rm eff})$ samples are needed in the approximation of Hessian matrices, where $d^\gamma_{\rm eff}$ is the $\gamma$-ridge leverage and can be much smaller than $d$ as long as $n\gamma \gg 1$. Additionally, we extend this result so that subsampled Newton methods can work for high-dimensional data on both distributed optimization problems and non-smooth regularized problems.

Click to Read Paper and Get Code
We study data-driven methods for community detection on graphs, an inverse problem that is typically solved in terms of the spectrum of certain operators or via posterior inference under certain probabilistic graphical models. Focusing on random graph families such as the stochastic block model, recent research has unified both approaches and identified both statistical and computational signal-to-noise detection thresholds. This graph inference task can be recast as a node-wise graph classification problem, and, as such, computational detection thresholds can be translated in terms of learning within appropriate models. We present a novel family of Graph Neural Networks (GNNs) and show that they can reach those detection thresholds in a purely data-driven manner without access to the underlying generative models, and even improve upon current computational thresholds in hard regimes. For that purpose, we propose to augment GNNs with the non-backtracking operator, defined on the line graph of edge adjacencies. We also perform the first analysis of optimization landscape on using GNNs to solve community detection problems, demonstrating that under certain simplifications and assumptions, the loss value at the local minima is close to the loss value at the global minimum/minima. Finally, the resulting model is also tested on real datasets, performing significantly better than previous models.

Click to Read Paper and Get Code
Single image super resolution is of great importance as a low-level computer vision task. Recent approaches with deep convolutional neural networks have achieved im-pressive performance. However, existing architectures have limitations due to the less sophisticated structure along with less strong representational power. In this work, to significantly enhance the feature representation, we proposed Triple Attention mixed link Network (TAN) which consists of 1) three different aspects (i.e., kernel, spatial and channel) of attention mechanisms and 2) fu-sion of both powerful residual and dense connections (i.e., mixed link). Specifically, the network with multi kernel learns multi hierarchical representations under different receptive fields. The output features are recalibrated by the effective kernel and channel attentions and feed into next layer partly residual and partly dense, which filters the information and enable the network to learn more powerful representations. The features finally pass through the spatial attention in the reconstruction network which generates a fusion of local and global information, let the network restore more details and improves the quality of reconstructed images. Thanks to the diverse feature recalibrations and the advanced information flow topology, our proposed model is strong enough to per-form against the state-of-the-art methods on the bench-mark evaluations.

Click to Read Paper and Get Code
The infinite restricted Boltzmann machine (iRBM) is an extension of the classic RBM. It enjoys a good property of automatically deciding the size of the hidden layer according to specific training data. With sufficient training, the iRBM can achieve a competitive performance with that of the classic RBM. However, the convergence of learning the iRBM is slow, due to the fact that the iRBM is sensitive to the ordering of its hidden units, the learned filters change slowly from the left-most hidden unit to right. To break this dependency between neighboring hidden units and speed up the convergence of training, a novel training strategy is proposed. The key idea of the proposed training strategy is randomly regrouping the hidden units before each gradient descent step. Potentially, a mixing of infinite many iRBMs with different permutations of the hidden units can be achieved by this learning method, which has a similar effect of preventing the model from over-fitting as the dropout. The original iRBM is also modified to be capable of carrying out discriminative training. To evaluate the impact of our method on convergence speed of learning and the model's generalization ability, several experiments have been performed on the binarized MNIST and CalTech101 Silhouettes datasets. Experimental results indicate that the proposed training strategy can greatly accelerate learning and enhance generalization ability of iRBMs.

* Submitted to Machine Learning
Click to Read Paper and Get Code
Recent work in learning ontologies (hierarchical and partially-ordered structures) has leveraged the intrinsic geometry of spaces of learned representations to make predictions that automatically obey complex structural constraints. We explore two extensions of one such model, the order-embedding model for hierarchical relation learning, with an aim towards improved performance on text data for commonsense knowledge representation. Our first model jointly learns ordering relations and non-hierarchical knowledge in the form of raw text. Our second extension exploits the partial order structure of the training data to find long-distance triplet constraints among embeddings which are poorly enforced by the pairwise training procedure. We find that both incorporating free text and augmented training constraints improve over the original order-embedding model and other strong baselines.

* 4 pages, ICML 2017 DeepStruct Workshop
Click to Read Paper and Get Code
Zero-shot learning (ZSL) models rely on learning a joint embedding space where both textual/semantic description of object classes and visual representation of object images can be projected to for nearest neighbour search. Despite the success of deep neural networks that learn an end-to-end model between text and images in other vision problems such as image captioning, very few deep ZSL model exists and they show little advantage over ZSL models that utilise deep feature representations but do not learn an end-to-end embedding. In this paper we argue that the key to make deep ZSL models succeed is to choose the right embedding space. Instead of embedding into a semantic space or an intermediate space, we propose to use the visual space as the embedding space. This is because that in this space, the subsequent nearest neighbour search would suffer much less from the hubness problem and thus become more effective. This model design also provides a natural mechanism for multiple semantic modalities (e.g., attributes and sentence descriptions) to be fused and optimised jointly in an end-to-end manner. Extensive experiments on four benchmarks show that our model significantly outperforms the existing models.

* To appear in CVPR 2017
Click to Read Paper and Get Code
Most existing person re-identification (re-id) methods focus on learning the optimal distance metrics across camera views. Typically a person's appearance is represented using features of thousands of dimensions, whilst only hundreds of training samples are available due to the difficulties in collecting matched training images. With the number of training samples much smaller than the feature dimension, the existing methods thus face the classic small sample size (SSS) problem and have to resort to dimensionality reduction techniques and/or matrix regularisation, which lead to loss of discriminative power. In this work, we propose to overcome the SSS problem in re-id distance metric learning by matching people in a discriminative null space of the training data. In this null space, images of the same person are collapsed into a single point thus minimising the within-class scatter to the extreme and maximising the relative between-class separation simultaneously. Importantly, it has a fixed dimension, a closed-form solution and is very efficient to compute. Extensive experiments carried out on five person re-identification benchmarks including VIPeR, PRID2011, CUHK01, CUHK03 and Market1501 show that such a simple approach beats the state-of-the-art alternatives, often by a big margin.

* accepted by CVPR2016
Click to Read Paper and Get Code
Extracting multi-scale information is key to semantic segmentation. However, the classic convolutional neural networks (CNNs) encounter difficulties in achieving multi-scale information extraction: expanding convolutional kernel incurs the high computational cost and using maximum pooling sacrifices image information. The recently developed dilated convolution solves these problems, but with the limitation that the dilation rates are fixed and therefore the receptive field cannot fit for all objects with different sizes in the image. We propose an adaptivescale convolutional neural network (ASCNet), which introduces a 3-layer convolution structure in the end-to-end training, to adaptively learn an appropriate dilation rate for each pixel in the image. Such pixel-level dilation rates produce optimal receptive fields so that the information of objects with different sizes can be extracted at the corresponding scale. We compare the segmentation results using the classic CNN, the dilated CNN and the proposed ASCNet on two types of medical images (The Herlev dataset and SCD RBC dataset). The experimental results show that ASCNet achieves the highest accuracy. Moreover, the automatically generated dilation rates are positively correlated to the sizes of the objects, confirming the effectiveness of the proposed method.

Click to Read Paper and Get Code
Treebanks traditionally treat punctuation marks as ordinary words, but linguists have suggested that a tree's "true" punctuation marks are not observed (Nunberg, 1990). These latent "underlying" marks serve to delimit or separate constituents in the syntax tree. When the tree's yield is rendered as a written sentence, a string rewriting mechanism transduces the underlying marks into "surface" marks, which are part of the observed (surface) string but should not be regarded as part of the tree. We formalize this idea in a generative model of punctuation that admits efficient dynamic programming. We train it without observing the underlying marks, by locally maximizing the incomplete data likelihood (similarly to EM). When we use the trained model to reconstruct the tree's underlying punctuation, the results appear plausible across 5 languages, and in particular, are consistent with Nunberg's analysis of English. We show that our generative model can be used to beat baselines on punctuation restoration. Also, our reconstruction of a sentence's underlying punctuation lets us appropriately render the surface punctuation (via our trained underlying-to-surface mechanism) when we syntactically transform the sentence.

Click to Read Paper and Get Code
Variance reduction techniques like SVRG provide simple and fast algorithms for optimizing a convex finite-sum objective. For nonconvex objectives, these techniques can also find a first-order stationary point (with small gradient). However, in nonconvex optimization it is often crucial to find a second-order stationary point (with small gradient and almost PSD hessian). In this paper, we show that Stabilized SVRG (a simple variant of SVRG) can find an $\epsilon$-second-order stationary point using only $\widetilde{O}(n^{2/3}/\epsilon^2+n/\epsilon^{1.5})$ stochastic gradients. To our best knowledge, this is the first second-order guarantee for a simple variant of SVRG. The running time almost matches the known guarantees for finding $\epsilon$-first-order stationary points.

Click to Read Paper and Get Code
In this article, we address the problem of computing adaptive sensorimotor models that can be used for guiding the motion of robotic systems with uncertain action-to-perception relations. The formulation of the uncalibrated sensor-based control problem is first presented, then, various computational methods for building adaptive sensorimotor models are derived and analysed. The proposed methodology is exemplified with two cases of study: (i) shape control of deformable objects with unknown properties, and (ii) soft manipulation of ultrasonic probes with uncalibrated sensors.

* 38th Chinese Control Conference
Click to Read Paper and Get Code