We generalized a modified exponentialized estimator by pushing the robust-optimal (RO) index $\lambda$ to $-\infty$ for achieving robustness to outliers by optimizing a quasi-Minimin function. The robustness is realized and controlled adaptively by the RO index without any predefined threshold. Optimality is guaranteed by expansion of the convexity region in the Hessian matrix to largely avoid local optima. Detailed quantitative analysis on both robustness and optimality are provided. The results of proposed experiments on fitting tasks for three noisy non-convex functions and the digits recognition task on the MNIST dataset consolidate the conclusions.

Continual learning in task-oriented dialogue systems can allow us to add new domains and functionalities through time without incurring the high cost of a whole system retraining. In this paper, we propose a continual learning benchmark for task-oriented dialogue systems with 37 domains to be learned continuously in four settings, such as intent recognition, state tracking, natural language generation, and end-to-end. Moreover, we implement and compare multiple existing continual learning baselines, and we propose a simple yet effective architectural method based on residual adapters. Our experiments demonstrate that the proposed architectural method and a simple replay-based strategy perform comparably well but they both achieve inferior performance to the multi-task learning baseline, in where all the data are shown at once, showing that continual learning in task-oriented dialogue systems is a challenging task. Furthermore, we reveal several trade-offs between different continual learning methods in term of parameter usage and memory size, which are important in the design of a task-oriented dialogue system. The proposed benchmark is released together with several baselines to promote more research in this direction.

Click through rate (CTR) prediction is very important for Native advertisement but also hard as there is no direct query intent. In this paper we propose a large-scale event embedding scheme to encode the each user browsing event by training a Siamese network with weak supervision on the users' consecutive events. The CTR prediction problem is modeled as a supervised recurrent neural network, which naturally model the user history as a sequence of events. Our proposed recurrent models utilizing pretrained event embedding vectors and an attention layer to model the user history. Our experiments demonstrate that our model significantly outperforms the baseline and some variants.

We proposed a deep learning method for interpretable diabetic retinopathy (DR) detection. The visual-interpretable feature of the proposed method is achieved by adding the regression activation map (RAM) after the global averaging pooling layer of the convolutional networks (CNN). With RAM, the proposed model can localize the discriminative regions of an retina image to show the specific region of interest in terms of its severity level. We believe this advantage of the proposed deep learning model is highly desired for DR detection because in practice, users are not only interested with high prediction performance, but also keen to understand the insights of DR detection and why the adopted learning model works. In the experiments conducted on a large scale of retina image dataset, we show that the proposed CNN model can achieve high performance on DR detection compared with the state-of-the-art while achieving the merits of providing the RAM to highlight the salient regions of the input image.

Time series and signals are attracting more attention across statistics, machine learning and pattern recognition as it appears widely in the industry， especially in sensor and IoT related research and applications, but few advances has been achieved in effective time series visual analytics and interaction due to its temporal dimensionality and complex dynamics. Inspired by recent effort on using network metrics to characterize time series for classification, we present an approach to visualize time series as complex networks based on the first order Markov process in its temporal ordering. In contrast to the classical bar charts, line plots and other statistics based graph, our approach delivers more intuitive visualization that better preserves both the temporal dependency and frequency structures. It provides a natural inverse operation to map the graph back to raw signals, making it possible to use graph statistics to characterize time series for better visual exploration and statistical analysis. Our experimental results suggest the effectiveness on various tasks such as pattern discovery and classification on both synthetic and the real time series and sensor data.

This paper proposes a set of new error criteria and learning approaches, Adaptive Normalized Risk-Averting Training (ANRAT), to attack the non-convex optimization problem in training deep neural networks (DNNs). Theoretically, we demonstrate its effectiveness on global and local convexity lower-bounded by the standard $L_p$-norm error. By analyzing the gradient on the convexity index $\lambda$, we explain the reason why to learn $\lambda$ adaptively using gradient descent works. In practice, we show how this method improves training of deep neural networks to solve visual recognition tasks on the MNIST and CIFAR-10 datasets. Without using pretraining or other tricks, we obtain results comparable or superior to those reported in recent literature on the same tasks using standard ConvNets + MSE/cross entropy. Performance on deep/shallow multilayer perceptrons and Denoised Auto-encoders is also explored. ANRAT can be combined with other quasi-Newton training methods, innovative network variants, regularization techniques and other specific tricks in DNNs. Other than unsupervised pretraining, it provides a new perspective to address the non-convex optimization problem in DNNs.

This paper proposes a set of new error criteria and a learning approach, called Adaptive Normalized Risk-Averting Training (ANRAT) to attack the non-convex optimization problem in training deep neural networks without pretraining. Theoretically, we demonstrate its effectiveness based on the expansion of the convexity region. By analyzing the gradient on the convexity index $\lambda$, we explain the reason why our learning method using gradient descent works. In practice, we show how this training method is successfully applied for improved training of deep neural networks to solve visual recognition tasks on the MNIST and CIFAR-10 datasets. Using simple experimental settings without pretraining and other tricks, we obtain results comparable or superior to those reported in recent literature on the same tasks using standard ConvNets + MSE/cross entropy. Performance on deep/shallow multilayer perceptron and Denoised Auto-encoder is also explored. ANRAT can be combined with other quasi-Newton training methods, innovative network variants, regularization techniques and other common tricks in DNNs. Other than unsupervised pretraining, it provides a new perspective to address the non-convex optimization strategy in training DNNs.

Inspired by recent successes of deep learning in computer vision, we propose a novel framework for encoding time series as different types of images, namely, Gramian Angular Summation/Difference Fields (GASF/GADF) and Markov Transition Fields (MTF). This enables the use of techniques from computer vision for time series classification and imputation. We used Tiled Convolutional Neural Networks (tiled CNNs) on 20 standard datasets to learn high-level features from the individual and compound GASF-GADF-MTF images. Our approaches achieve highly competitive results when compared to nine of the current best time series classification approaches. Inspired by the bijection property of GASF on 0/1 rescaled data, we train Denoised Auto-encoders (DA) on the GASF images of four standard and one synthesized compound dataset. The imputation MSE on test data is reduced by 12.18%-48.02% when compared to using the raw data. An analysis of the features and weights learned via tiled CNNs and DAs explains why the approaches work.

Inspired by recent successes of deep learning in computer vision and speech recognition, we propose a novel framework to encode time series data as different types of images, namely, Gramian Angular Fields (GAF) and Markov Transition Fields (MTF). This enables the use of techniques from computer vision for classification. Using a polar coordinate system, GAF images are represented as a Gramian matrix where each element is the trigonometric sum (i.e., superposition of directions) between different time intervals. MTF images represent the first order Markov transition probability along one dimension and temporal dependency along the other. We used Tiled Convolutional Neural Networks (tiled CNNs) on 12 standard datasets to learn high-level features from individual GAF, MTF, and GAF-MTF images that resulted from combining GAF and MTF representations into a single image. The classification results of our approach are competitive with five stateof-the-art approaches. An analysis of the features and weights learned via tiled CNNs explains why the approach works.