Learning Convolutional Networks for Content-weighted Image Compression

Lossy image compression is generally formulated as a joint rate-distortion optimization to learn encoder, quantizer, and decoder. However, the quantizer is non-differentiable, and discrete entropy estimation usually is required for rate control. These make it very challenging to develop a convolutional network (CNN)-based image compression system. In this paper, motivated by that the local information content is spatially variant in an image, we suggest that the bit rate of the different parts of the image should be adapted to local content. And the content aware bit rate is allocated under the guidance of a content-weighted importance map. Thus, the sum of the importance map can serve as a continuous alternative of discrete entropy estimation to control compression rate. And binarizer is adopted to quantize the output of encoder due to the binarization scheme is also directly defined by the importance map. Furthermore, a proxy function is introduced for binary operation in backward propagation to make it differentiable. Therefore, the encoder, decoder, binarizer and importance map can be jointly optimized in an end-to-end manner by using a subset of the ImageNet database. In low bit rate image compression, experiments show that our system significantly outperforms JPEG and JPEG 2000 by structural similarity (SSIM) index, and can produce the much better visual result with sharp edges, rich textures, and fewer artifacts.

Learning Domain-Invariant Subspace using Domain Features and Independence Maximization

Domain adaptation algorithms are useful when the distributions of the training and the test data are different. In this paper, we focus on the problem of instrumental variation and time-varying drift in the field of sensors and measurement, which can be viewed as discrete and continuous distributional change in the feature space. We propose maximum independence domain adaptation (MIDA) and semi-supervised MIDA (SMIDA) to address this problem. Domain features are first defined to describe the background information of a sample, such as the device label and acquisition time. Then, MIDA learns a subspace which has maximum independence with the domain features, so as to reduce the inter-domain discrepancy in distributions. A feature augmentation strategy is also designed to project samples according to their backgrounds so as to improve the adaptation. The proposed algorithms are flexible and fast. Their effectiveness is verified by experiments on synthetic datasets and four real-world ones on sensors, measurement, and computer vision. They can greatly enhance the practicability of sensor systems, as well as extend the application scope of existing domain adaptation algorithms by uniformly handling different kinds of distributional change.

Multi-channel Weighted Nuclear Norm Minimization for Real Color Image Denoising

Most of the existing denoising algorithms are developed for grayscale images, while it is not a trivial work to extend them for color image denoising because the noise statistics in R, G, B channels can be very different for real noisy images. In this paper, we propose a multi-channel (MC) optimization model for real color image denoising under the weighted nuclear norm minimization (WNNM) framework. We concatenate the RGB patches to make use of the channel redundancy, and introduce a weight matrix to balance the data fidelity of the three channels in consideration of their different noise statistics. The proposed MC-WNNM model does not have an analytical solution. We reformulate it into a linear equality-constrained problem and solve it with the alternating direction method of multipliers. Each alternative updating step has closed-form solution and the convergence can be guaranteed. Extensive experiments on both synthetic and real noisy image datasets demonstrate the superiority of the proposed MC-WNNM over state-of-the-art denoising methods.

GPU Activity Prediction using Representation Learning

GPU activity prediction is an important and complex problem. This is due to the high level of contention among thousands of parallel threads. This problem was mostly addressed using heuristics. We propose a representation learning approach to address this problem. We model any performance metric as a temporal function of the executed instructions with the intuition that the flow of instructions can be identified as distinct activities of the code. Our experiments show high accuracy and non-trivial predictive power of representation learning on a benchmark.

Low Precision Neural Networks using Subband Decomposition

Large-scale deep neural networks (DNN) have been successfully used in a number of tasks from image recognition to natural language processing. They are trained using large training sets on large models, making them computationally and memory intensive. As such, there is much interest in research development for faster training and test time. In this paper, we present a unique approach using lower precision weights for more efficient and faster training phase. We separate imagery into different frequency bands (e.g. with different information content) such that the neural net can better learn using less bits. We present this approach as a complement existing methods such as pruning network connections and encoding learning weights. We show results where this approach supports more stable learning with 2-4X reduction in precision with 17X reduction in DNN parameters.

Deep Identity-aware Transfer of Facial Attributes

This paper presents a Deep convolutional network model for Identity-Aware Transfer (DIAT) of facial attributes. Given the source input image and the reference attribute, DIAT aims to generate a facial image (i.e., target image) that not only owns the reference attribute but also keep the same or similar identity to the input image. We develop a two-stage scheme to transfer the input image to each reference attribute label. A feed-forward transform network is first trained by combining perceptual identity-aware loss and GAN-based attribute loss, and a face enhancement network is then introduced to improve the visual quality. We further define perceptual identity loss on the convolutional feature maps of the attribute discriminator, resulting in a DIAT-A model. Our DIAT and DIAT-A models can provide a unified solution for several representative facial attribute transfer tasks such as expression transfer, accessory removal, age progression, and gender transfer. The experimental results validate their effectiveness. Even for some identity-related attribute (e.g., gender), our DIAT-A can obtain visually impressive results by changing the attribute while retaining most identity features of the source image.

Evolutionary Cost-sensitive Extreme Learning Machine

Conventional extreme learning machines solve a Moore-Penrose generalized inverse of hidden layer activated matrix and analytically determine the output weights to achieve generalized performance, by assuming the same loss from different types of misclassification. The assumption may not hold in cost-sensitive recognition tasks, such as face recognition based access control system, where misclassifying a stranger as a family member may result in more serious disaster than misclassifying a family member as a stranger. Though recent cost-sensitive learning can reduce the total loss with a given cost matrix that quantifies how severe one type of mistake against another, in many realistic cases the cost matrix is unknown to users. Motivated by these concerns, this paper proposes an evolutionary cost-sensitive extreme learning machine (ECSELM), with the following merits: 1) to our best knowledge, it is the first proposal of ELM in evolutionary cost-sensitive classification scenario; 2) it well addresses the open issue of how to define the cost matrix in cost-sensitive learning tasks; 3) an evolutionary backtracking search algorithm is induced for adaptive cost matrix optimization. Experiments in a variety of cost-sensitive tasks well demonstrate the effectiveness of the proposed approaches, with about 5%~10% improvements.

Convolutional Network for Attribute-driven and Identity-preserving Human Face Generation

This paper focuses on the problem of generating human face pictures from specific attributes. The existing CNN-based face generation models, however, either ignore the identity of the generated face or fail to preserve the identity of the reference face image. Here we address this problem from the view of optimization, and suggest an optimization model to generate human face with the given attributes while keeping the identity of the reference image. The attributes can be obtained from the attribute-guided image or by tuning the attribute features of the reference image. With the deep convolutional network "VGG-Face", the loss is defined on the convolutional feature maps. We then apply the gradient decent algorithm to solve this optimization problem. The results validate the effectiveness of our method for attribute driven and identity-preserving face generation.

Robust Visual Knowledge Transfer via EDA

We address the problem of visual knowledge adaptation by leveraging labeled patterns from source domain and a very limited number of labeled instances in target domain to learn a robust classifier for visual categorization. This paper proposes a new extreme learning machine based cross-domain network learning framework, that is called Extreme Learning Machine (ELM) based Domain Adaptation (EDA). It allows us to learn a category transformation and an ELM classifier with random projection by minimizing the l_(2,1)-norm of the network output weights and the learning error simultaneously. The unlabeled target data, as useful knowledge, is also integrated as a fidelity term to guarantee the stability during cross domain learning. It minimizes the matching error between the learned classifier and a base classifier, such that many existing classifiers can be readily incorporated as base classifiers. The network output weights cannot only be analytically determined, but also transferrable. Additionally, a manifold regularization with Laplacian graph is incorporated, such that it is beneficial to semi-supervised learning. Extensively, we also propose a model of multiple views, referred as MvEDA. Experiments on benchmark visual datasets for video event recognition and object recognition, demonstrate that our EDA methods outperform existing cross-domain learning methods.

Multi-modal Fusion for Diabetes Mellitus and Impaired Glucose Regulation Detection

Effective and accurate diagnosis of Diabetes Mellitus (DM), as well as its early stage Impaired Glucose Regulation (IGR), has attracted much attention recently. Traditional Chinese Medicine (TCM) [3], [5] etc. has proved that tongue, face and sublingual diagnosis as a noninvasive method is a reasonable way for disease detection. However, most previous works only focus on a single modality (tongue, face or sublingual) for diagnosis, although different modalities may provide complementary information for the diagnosis of DM and IGR. In this paper, we propose a novel multi-modal classification method to discriminate between DM (or IGR) and healthy controls. Specially, the tongue, facial and sublingual images are first collected by using a non-invasive capture device. The color, texture and geometry features of these three types of images are then extracted, respectively. Finally, our so-called multi-modal similar and specific learning (MMSSL) approach is proposed to combine features of tongue, face and sublingual, which not only exploits the correlation but also extracts individual components among them. Experimental results on a dataset consisting of 192 Healthy, 198 DM and 114 IGR samples (all samples were obtained from Guangdong Provincial Hospital of Traditional Chinese Medicine) substantiate the effectiveness and superiority of our proposed method for the diagnosis of DM and IGR, compared to the case of using a single modality.