This article proposes the theory and the architecture of Knowledge Programmable Intelligent Chip Systems(KPI-CS). KPI-CS is based on cutting-edge heterogeneous computing and reconfigurable AI chip technologies, fusing complex system computing, knowledge engineering and semiconductor IC design technology. It is aimed at providing adaptability to different application scenarios, flexibility in chip architecture reconfiguration and rationality in AI algorithmic computing capability to support parallel intelligent systems. KPI-CS can provide effective and efficient real-time supporting computing facilities which adapt to different demands in intelligent systems.

A computing method for matrix-weighted itemset support is proposed firstly, and the algorithm of matrix-weighted association patterns mining for cross-language query expansion is presented. Then, the algorithm of cross-language query post-translation expansion is put forward based on matrix-weighted association rules mining. The first cross-language retrieval is performed to obtain the top initially retrieved documents(TIRDs) by machine translation, and the relevance feedback documents(RFDs) are gained from TIRDs by user correlation judgment. The matrix-weighted frequent itemsets containing original query terms are mined from RFDs by means of computing support and the association rules with original query terms are extracted from frequent itemsets according to the evaluation framework of confidence-interest. To implement cross-language query post-translation expansion, the consequents or antecedents of the rules are treated as expansion terms and the importance of the expansion terms is measured by the confidence and interest of the rule. Experiments on NTCIR-5 CLIR standard test set show that the proposed algorithm improves the performance of cross-language query expansion, and it is beneficial in cross-language retrieval of long queries. The performance of post-translation consequent expansion is better than that of the antecedent one.

The existing personalized learning resource recommendation methods have problems of single model and low speed and matching degree. Therefore a personalized learning resource recommendation method based on stage evolution bidirectional self-balancing(EBPLRM-M) is proposed. Firstly, a personalized learning resource recommendation model is built based on stage evolution bidirectional self-balancing. The parameterized representations of resource recommendation features and fitness functions are improved. Then, a fuzzy adaptive binary particle swarm optimization algorithm based on evolutionary state determination is adopted to optimize the model. Finally, simulation results show that EBPLRM-M has higher matching degree and recommended speed compared with the recommendation methods adopting the classical algorithms.

Segmenting the plant leaf by a fixed model is problematic due to different imaging conditions, such as backgrounds, shadows and reflections. In this paper,interactive plant leaf segmentation method based on robust random walk is proposed. The interactive strategy is employed to propagate prior information of user-specified pixel, and robust random walk algorithm is used to realize plant leaf segmentation. The relationship of pairwise pixels is considered based on random walk and a super-pixel-consistent constraint is added to make the edges of segmentation smooth. Since the random walk algorithm only considers pairwise pixel relations, it is sensitive to the specified and connected pixels. The prior information of specified pixel is obtained through human-machine interaction. On the basis of random walk model, a log-likelihood ratio is used to predict the probability of a pixel belonging to the background and guide the label propagation. Experiments on plant leaf images in loose controlled and uncontrolled environments show that the proposed method obtains smoother and more robust plant leaf segmentation images than other methods.

An exemplars multiple kernel learning-extreme learning machine(MKL-ELM) based human action recognition approach with multi-modal visual feature fusion from RGB-D videos is proposed to solve the problem of single modal visual feature with the limited discrimination ability for all categories of human actions. Firstly, the robust and dense moving pose features with human skeleton surface fitting and dense trajectories from human motion are extracted. The sparse histogram of oriented principal component(SHOPC) features of 3D body geometry with the normal plane of dense point clouds is perceived and the histogram of 3D gradient orientation(HOG3D) features embedded with human appearance textures on spatial temporal neighbor of body nodes in the given videos is extracted. The modified MKL-ELM with radius-margin bound is exploited to fuse the given multi-modal visual features. Then, the set of the representative exemplars for each human action is mined with the contrast data technique. Finally, each sample is hierarchically classified by the designed exemplars-MKL-ELM model with greedy prediction strategy to recognize the human actions with the fused features and the given exemplars. The experiments show that compared with the traditional methods, the proposed action recognition method has significant advantages with high classification accuracy and computational efficiency.