In this paper, we study the topical behavior in a large scale. We use the network logs where each entry contains the entity ID, the timestamp, and the meta data about the activity. Both the temporal and the spatial relationships of the behavior are explored with the deep learning architectures combing the recurrent neural network (RNN) and the convolutional neural network (CNN). To make the behavioral data appropriate for the spatial learning in the CNN, we propose several reduction steps to form the topical metrics and to place them homogeneously like pixels in the images. The experimental result shows both temporal and spatial gains when compared against a multilayer perceptron (MLP) network. A new learning framework called the spatially connected convolutional networks (SCCN) is introduced to predict the topical metrics more efficiently.

In this paper we present a new evolutionary algorithm for the protein folding problem. We study the problem in the 3D FCC HP model which has been widely used in previous research. Our focus is to develop evolutionary algorithms (EA) which are robust, easy to operate and can handle various energy functions. We propose lattice rotation for crossover and K-site move for mutation, which form the key components of our evolutionary algorithms. Experiment shows that our algorithms are able to find minimum-energy conformations for many sequences whose optimal conformations are not found in previous EA-based algorithms. Furthermore, our idea can be easily integrated into Monte Carlo and Tabu searches as approaches for local searches.

In this article we study layout and circuit implementations of 3-input lookup table (3-LUT) for via configurable structured ASIC. We present a new 3-LUT circuit and several layout designs. We also propose a method to improve the delay of any logic function with fewer inputs. A 3-LUT, being able to realize all the 256 3-input functions, enables us to synthesize a circuit using both a standard cell synthesizer and an FPGA technology mapper such as FlowMap. Our study shows that circuits synthesized using a standard-cell synthesizer usually achieves better timing than that obtained by FlowMap. Our study further shows that the well-known 3-LUT implemented with multiplexers achieves better timing, area, and power dissipation. Our methodology can be also employed to study look-up tables with more inputs.

MOTIVATION: Missing data in genotyping single nucleotide polymorphism (SNP) spots are common. High-throughput genotyping methods usually have a high rate of missing data. For example, the published human chromosome 21 data by Patil et al. contains about 20% missing SNPs. Inferring missing SNPs using the haplotype block structure is promising but difficult because the haplotype block boundaries are not well defined. Here we propose a global algorithm to overcome this difficulty. RESULTS: First, we propose to use entropy as a measure of haplotype diversity. We show that the entropy measure combined with a dynamic programming algorithm produces better haplotype block partitions than other measures. Second, based on the entropy measure, we propose a two-step iterative partition-inference algorithm for the inference of missing SNPs. At the first step, we apply the dynamic programming algorithm to partition haplotypes into blocks. At the second step, we use an iterative process similar to the expectation-maximization algorithm to infer missing SNPs in each haplotype block so as to minimize the block entropy. The algorithm iterates these two steps until the total block entropy is minimized. We test our algorithm in several experimental data sets. The results show that the global approach significantly improves the accuracy of the inference. AVAILABILITY: Upon request.