Timely accurate traffic forecast is crucial for urban control and guidance. Due to the high nonlinearity complexity of flow, traditional methods cannot satisfy requirements mid-and-long term prediction tasks often neglect spatial temporal dependencies. In this paper, we propose a novel deep learning framework, Spatio-Temporal Graph Convolutional Networks (STGCN), tackle time series problem in domain. Instead applying regular convolutional recurrent units, formulate on graphs build model with...

Spatial-temporal data forecasting of traffic flow is a challenging task because complicated spatial dependencies and dynamical trends temporal pattern between different roads. Existing frameworks usually utilize given adjacency graph sophisticated mechanisms for modeling correlations. However, limited representations structure with incomplete adjacent connections may restrict effective spatial-temporal learning those models. Furthermore, existing methods were out at elbows when solving data:...

Graph Convolutional Networks (GCNs) play a crucial role in graph learning tasks, however, embedding with few supervised signals is still difficult problem. In this paper, we propose novel training algorithm for Network, called Multi-Stage Self-Supervised (M3S) Training Algorithm, combined self-supervised approach, focusing on improving the generalization performance of GCNs graphs labeled nodes. Firstly, Framework provided as basis M3S method. Then leverage DeepCluster technique, popular...

Most artificial intelligence models have limiting ability to solve new tasks faster, without forgetting previously acquired knowledge. The recently emerging paradigm of continual learning aims this issue, in which the model learns various a sequential fashion. In work, novel approach for is proposed, searches best neural architecture each coming task via sophisticatedly designed reinforcement strategies. We name it as Reinforced Continual Learning. Our method not only has good performance on...

It is widely observed that deep learning models with learned parameters generalize well, even much more model than the number of training samples. We systematically investigate underlying reasons why neural networks often and reveal difference between minima (with same error) well those they don't. show it characteristics landscape loss function explains good generalization capability. For for networks, volume basin attraction dominates over poor minima, which guarantees optimization methods...

Despite several attacks have been proposed, text-based CAPTCHAs are still being widely used as a security mechanism. One of the reasons for pervasive use text captchas is that many prior scheme-specific and require labor-intensive time-consuming process to construct. This means change in captcha features like noisier background can simply invalid an earlier attack. paper presents generic, yet effective solver based on generative adversarial network. Unlike machine-learning-based approaches...

Neural architecture search (NAS) attracts much research attention because of its ability to identify better architectures than handcrafted ones. Recently, differentiable methods become the state-of-the-arts on NAS, which can obtain high-performance in several days. However, they still suffer from huge computation costs and inferior performance due construction supernet. In this paper, we propose an efficient NAS method based proximal iterations (denoted as NASP). Different previous works,...

Deep learning achieves state-of-the-art results in many tasks computer vision and natural language processing. However, recent works have shown that deep networks can be vulnerable to adversarial perturbations, which raised a serious robustness issue of networks. Adversarial training, typically formulated as robust optimization problem, is an effective way improving the A major drawback existing training algorithms computational overhead generation examples, far greater than network...

State-of-the-art deep neural networks are known to be vulnerable adversarial examples, formed by applying small but malicious perturbations the original inputs. Moreover, can \textit{transfer across models}: examples generated for a specific model will often mislead other unseen models. Consequently adversary leverage it attack deployed systems without any query, which severely hinder application of learning, especially in areas where security is crucial. In this work, we systematically...

Bingfeng Luo, Yansong Feng, Zheng Wang, Zhanxing Zhu, Songfang Huang, Rui Yan, Dongyan Zhao. Proceedings of the 55th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2017.

Data-driven modeling of human motions is ubiquitous in computer graphics and vision applications, such as synthesizing realistic or recognizing actions. Recent research has shown that problems can be approached by learning a natural motion manifold using deep on large amount data, to address the shortcomings traditional data-driven approaches. However, previous methods sub-optimal for two reasons. First, skeletal information not been fully utilized feature extraction. Unlike images, it...

The spatio-temporal graph learning is becoming an increasingly important object of study. Many application domains involve highly dynamic graphs where temporal information crucial, e.g. traffic networks and financial transaction graphs. Despite the constant progress made on structured data, there still a lack effective means to extract complex features from structures. Particularly, conventional models such as convolutional or recurrent neural are incapable revealing patterns in short long...

Convolutional Neural Networks (CNNs) are known to rely more on local texture rather than global shape when making decisions. Recent work also indicates a close relationship between CNN's texture-bias and its robustness against distribution shift, adversarial perturbation, random corruption, etc. In this work, we attempt at improving various kinds of universally by alleviating bias. With inspiration from the human visual system, propose light-weight model-agnostic method, namely Informative...

Spatio-temporal prediction plays an important role in many application areas especially traffic domain. However, due to complicated spatio-temporal dependency and high non-linear dynamics road networks, task is still challenging. Existing works either exhibit heavy training cost or fail accurately capture the patterns, also ignore correlation between distant roads that share similar patterns. In this paper, we propose a novel deep learning framework overcome these issues: 3D Temporal Graph...

Understanding the behavior of stochastic gradient descent (SGD) in context deep neural networks has raised lots concerns recently. Along this line, we study a general form based optimization dynamics with unbiased noise, which unifies SGD and standard Langevin dynamics. Through investigating dynamics, analyze on escaping from minima its regularization effects. A novel indicator is derived to characterize efficiency through measuring alignment noise covariance curvature loss function. Based...