Novel Low-Power Construction of Chaotic S-Box in Multilayer Perceptron.

Multilayer perceptron is composed of massive distributed neural processors interconnected. The nonlinear dynamic components in these processors expand the input data into a linear combination of synapses. However, the nonlinear mapping ability of original multilayer perceptron is limited when processing high complexity information. The introduction of more powerful nonlinear components (e.g., S-box) to multilayer perceptron can not only reinforce its information processing ability, but also enhance the overall security. Therefore, we combine the methods of cryptography and information theory to design a low-power chaotic S-box (LPC S-box) with entropy coding in the hidden layer to make the multilayer perceptron process information more efficiently and safely. In the performance test, our S-box architecture has good properties, which can effectively resist main known attacks (e.g., Berlekamp Massey-attack and Ronjom-Helleseth attack). This interdisciplinary work can attract more attention from academia and industry to the security of multilayer perceptron.

Temporal event searches based on event maps and relationships.

To satisfy a user's need to find and understand the whole picture of an event effectively and efficiently, in this paper we formalize the problem of temporal event searches and propose a framework of event relationship analysis for search events based on user queries. We define three kinds of event relationships: temporal, content dependence, and event reference, that can be used to identify to what extent a component event is dependent on another in the evolution of a target event (i.e., the query event). The search results are organized as a temporal event map (TEM) that serves as the whole picture about an event's evolution or development by showing the dependence relationships among events. Based on the event relationships in the TEM, we further propose a method to measure the degrees of importance of events, so as to discover the important component events for a query, as well as the several algebraic operators involved in the TEM, that allow users to view the target event. Experiments conducted on a real data set show that our method outperforms the baseline method Event Evolution Graph (EEG), and it can help discover certain new relationships missed by previous methods and even by human annotators.

Anchor Link Prediction across Attributed Networks via Network Embedding.

Presently, many users are involved in multiple social networks. Identifying the same user in different networks, also known as anchor link prediction, becomes an important problem, which can serve numerous applications, e.g., cross-network recommendation, user profiling, etc. Previous studies mainly use hand-crafted structure features, which, if not carefully designed, may fail to reflect the intrinsic structure regularities. Moreover, most of the methods neglect the attribute information of social networks. In this paper, we propose a novel semi-supervised network-embedding model to address the problem. In the model, each node of the multiple networks is represented by a vector for anchor link prediction, which is learnt with awareness of observed anchor links as semi-supervised information, and topology structure and attributes as input. Experimental results on the real-world data sets demonstrate the superiority of the proposed model compared to state-of-the-art techniques.

On the Effectiveness of Least Squares Generative Adversarial Networks.

Unsupervised learning with generative adversarial networks (GANs) has proven to be hugely successful. Regular GANs hypothesize the discriminator as a classifier with the sigmoid cross entropy loss function. However, we found that this loss function may lead to the vanishing gradients problem during the learning process. To overcome such a problem, we propose in this paper the Least Squares Generative Adversarial Networks (LSGANs) which adopt the least squares loss for both the discriminator and the generator. We show that minimizing the objective function of LSGAN yields minimizing the Pearson χ2 divergence. We also show that the derived objective function that yields minimizing the Pearson χ2 divergence performs better than the classical one of using least squares for classification. There are two benefits of LSGANs over regular GANs. First, LSGANs are able to generate higher quality images than regular GANs. Second, LSGANs perform more stably during the learning process. For evaluating the image quality, we conduct both qualitative and quantitative experiments, and the experimental results show that LSGANs can generate higher quality images than regular GANs. Furthermore, we evaluate the stability of LSGANs in two groups. One is to compare between LSGANs and regular GANs without gradient penalty. We conduct three experiments, including Gaussian mixture distribution, difficult architectures, and a newly proposed method - datasets with small variability, to illustrate the stability of LSGANs. The other one is to compare between LSGANs with gradient penalty (LSGANs-GP) and WGANs with gradient penalty (WGANs-GP). The experimental results show that LSGANs-GP succeed in training for all the difficult architectures used in WGANs-GP, including 101-layer ResNet.