The performance of deep neural networks (DNNs) often depends on the design of their architectures. But designing a DNN with good performance is a difficult and knowledge-intensive process. In this paper, we propose a neural architecture search method based on improved brain storm optimization (BSO) algorithm to efficiently deal with image classification tasks. BSO successfully transposes the human brainstorming process to design of optimization algorithms, which typically uses grouping, substitution, and creation operators to generate as many solutions as possible to approach the global optimization of the problem generation by generation. However, the BSO algorithm using clustering methods for grouping increases the computational burden, so we use the BSO algorithm based on simple grouping methods to solve the optimal architecture of the neural architecture search (NAS). We also redesigned the search space and designed an efficient encoding strategy for each individual.

Model compression technology investigates the compression of deep neural networks by quantizing the full-precision weights of the network into low-bit ones, to achieve network acceleration. However, most of the existing quantization operations are calculated by simple thresholding operations, which will lead to serious precision loss. In this paper, we propose a new quantization framework combined with pruning, called Multiple Residual Quantization of Pruning (MRQP), to achieve higher precision quantization neural network (QNN). MRQP recursively performs quantization of the full-precision weights by combining the low-bit weights stem and residual parts many times, to minimize the error between the quantized weights and the full-precision weights, and to ensure higher precision quantization. At the same time, MRQP prunes some weights that have less impact on loss function to further reduce model size.

Machine learning (ML), as the most promising paradigm to discover deep knowledge from data, has been widely applied to practical applications, such as recommender systems, virtual reality, and semantic segmentation. However, building a high-quality ML system for given tasks requires expert knowledge and high computation cost. This poses a significant challenge to the further development of ML in large-scale practical applications. The automatic design of ML has become an increasingly popular research trend. At the same time, evolutionary computation (EC), as an excellent heuristic search technique, has been widely employed in ML optimization, so-called evolutionary machine learning (EML). In this paper, we offer a comprehensive review of the literature (more than 500 references) for EML methods. We first introduce the concepts related to ML and EC. After that, we propose a taxonomy criterion based on the ML and EC perspectives. The important research problems of EML, e.g., ML algorithms, solution representations, search paradigms, acceleration strategies and applications, are reviewed systematically. Lastly, we analyze EML limitations and discuss potential trends that are promising to address in the future.

Large-scale multi-objective optimization problems (LSMOPs) can lead to the conventional reproduction operator being inefficient for searching. Therefore, we propose a large-scale multi-objective brain storm optimization algorithm based on direction vectors and variance analysis (LMOBSO-DV) to enhance the efficiency of tackling LSMOPs. Specifically, we adopt brain storm optimization (BSO) algorithm using reference vectors to divide the population into subpopulations and guide the individuals i) in each subpopulation to search in promising directions and 2) between subpopulations to maintain diversity. We also design a new mutation operator. On a widely used LSMOPs test suites with 1000 decision variables, 2 objectives, and 3 objectives, we evaluate LMOBSO-DV's effectiveness in comparison to other several state-of-the-art algorithms. The results of the experiment show that our proposed approach, LMOBSO-DV, outperforms the other studied algorithms.

为了解决刑期预测任务准确率较差的问题,提出一种基于多通道分层注意力循环神经网络的司法案件刑期预测模型.该模型对传统的循环神经网络模型进行了改进,引入了BERT词嵌入、多通道模式和分层注意力机制,将刑期预测转化为文本分类问题.模型采用分层的双向循环神经网络对案件文本进行建模,并通过分层注意力机制在词语级和句子级两个层面捕获不同词语和句子的重要性,最终生成有效表征案件文本的多通道嵌入向量.实验结果表明:对比现有的基于深度学习的刑期预测模型,本文提出的模型具有更高的预测性能.

近年来,伴随着人工智能的发展及法院裁判文书的公开化,"智慧司法"、案例推荐成为热点问题.针对案例推荐中存在的推荐准确性差、传统知识图谱向量化表示精度不高等问题,提出基于知识图谱的案件推荐(Knowledge Graph based Case Recommendation,KGCR)模型.该模型以知识图谱为辅助信息,利用文本分类和信息抽取技术构建面向刑事案例的知识图谱,针对当事人的陈词供述,利用知识表示学习求解相似的案件,进一步实现法条推荐.针对TransH算法的负采样问题进行改进,提出FU-TransH算法模型.以公开的刑事判决书为数据集进行实验,实验结果表明,与相关的具有代表性的算法相比,该算法的推荐准确率更高.

For the conflicting problem between memory assumption and access speed caused by existing caching methods of report,a caching method is provided for structured data.The main characteristic of such method is caching the report data which has the structure of row and column to file as data blocks.The data to be cached as file is partitioned into indexing area and data area,then written to the file via a data blocking algorithm.Using such approach,the need data can be retrieved from the located block by index,and the data rapidly.Thereby,an optimized result can be obtained to the memory assumption and access speed problem.