Search

Convolutional neural networks have achieved amazing success in many areas in recent years, and VGG is a widely used convolutional neural network model. However, it has some limitations, such as a large number of parameters, which take up significant memory and thus restrict its application in resource-constrained scenarios such as mobile devices and embedded systems. In convolutional neural network models, the different number of convolutional layers can extract different granularity features, which represent the different levels of importance in the image recognition process. Here, we propose a new VGG architecture with different granularity feature combinations that combine different granularity features from block1, block2, block3, block4, and block5 in VGG. Each block is followed by a local fully connected layer to reduce the dimensionality of the coarse and fine features, and five different granularity features are combined as the input of the first global fully connected layer. By combining the features of different blocks, it can increase information flow from a lower layer directly to a fully connected layer and increase feature reuse without adding too many connections. The addition of five local fully connected layers means an increase in parameters, so we reduce the neuron number in two global fully connected layers to reduce the number of parameters. The well-known datasets CIFAR-10 and MNIST were used to evaluate the network's classification performance. The experimental results show that the proposed model achieves better training and testing performance than traditional VGGs and reduces the number of parameters.

Fuzzy knowledge representation and reasoning is one of the key research subjects of artificial intelligence with uncertainty. Currently, fuzzy Petri nets (FPNs) cannot describe fuzzy nondeterministic knowledge based on the variable fuzzy criterion clearly. To solve this problem, we present an extended FPN model based on the qualitative mapping and the qualitative criterion transformation. First, the mechanism and the principles of the variable fuzzy criterion under qualitative mapping are introduced. The important theoretical basis of qualitative mapping is detection and integration of perceptual attribute features. Under transition degree function and the qualitative criterion transformation, the qualitative mapping better reflects the cases in which the fuzzy membership criterion is variable. Second, we use the characteristics of the qualitative mapping to propose a structure of attribute fuzzy information granule. Based on the logic operation of fuzzy information granule, we redescribe the four fuzzy production rules which are commonly used, and obtain four kinds of fuzzy qualitative judgment rules. They can explicitly express the occurrence process of transition nodes of FPNs, which strengthens the uncertain knowledge representation ability of FPNs. The new method makes it easy to construct a learning mechanism based on cognition in FPNs.

Background: Identifying Drug-Target Interactions (DTIs) is a major challenge for current drug discovery and drug repositioning. Compared to traditional experimental approaches, in silico methods are fast and inexpensive. With the increase in open-access experimental data, numerous computational methods have been applied to predict DTIs. Methods: In this study, we propose an end-to-end learning model of Factorization Machine and Deep Neural Network (FM-DNN), which emphasizes both low-order (first or second order) and high-order (higher than second order) feature interactions without any feature engineering other than raw features. This approach combines the power of FM and DNN learning for feature learning in a new neural network architecture. Results: The experimental DTI basic features include drug characteristics (609), target characteristics (1819), plus drug ID, target ID, total 2430. We compare 8 models such as SVM, GBDT, WIDE-DEEP etc, the FM-DNN algorithm model obtains the best results of AUC(0.8866) and AUPR(0.8281). Conclusion: Feature engineering is a job that requires expert knowledge, it is often difficult and time-consuming to achieve good results. FM-DNN can auto learn a lower-order expression by FM and a high-order expression by DNN.FM-DNN model has outstanding advantages over other commonly used models.

With the continuous development of the Internet, social media based on short text has become popular. However, the sparsity and shortness of essays will restrict the accuracy of text classification. Therefore, based on the Bert model, we capture the mental feature of reviewers and apply them for short text classification to improve its classification accuracy. Specifically, we construct a model text at the language level and fine tune the model to better embed mental features. To verify the accuracy of this method, we compare a variety of machine learning methods, such as support vector machine, convolution neural networks, and recurrent neural networks. The results show the following: (1) Through feature comparison, it is found that mental features can significantly improve the accuracy of short text classification. (2) Combining mental features and text as input vectors can provide more classification accuracy than separating them as two independent vectors. (3) Through model comparison, it can be found that Bert model can integrate mental features and short text. Bert can better capture mental features to improve the accuracy of classification results. This will help to promote the development of short text classification.

Recent progress in biomedical image segmentation based on deep convolutional neural networks (CNNs) has drawn much attention. However, its vulnerability towards adversarial samples cannot be overlooked. This paper is the first one that discovers that all the CNN-based state-of-the-art biomedical image segmentation models are sensitive to adversarial perturbations. This limits the deployment of these methods in safety-critical biomedical fields. In this paper, we discover that global spatial dependencies and global contextual information in a biomedical image can be exploited to defend against adversarial attacks. To this end, non-local context encoder (NLCE) is proposed to model short- and long range spatial dependencies and encode global contexts for strengthening feature activations by channel-wise attention. The NLCE modules enhance the robustness and accuracy of the non-local context encoding network (NLCEN), which learns robust enhanced pyramid feature representations with NLCE modules, and then integrates the information across different levels. Experiments on both lung and skin lesion segmentation datasets have demonstrated that NLCEN outperforms any other state-of-the-art biomedical image segmentation methods against adversarial attacks. In addition, NLCE modules can be applied to improve the robustness of other CNN-based biomedical image segmentation methods., Comment: Accepted by AAAI2019 as oral presentation

Theory of Inventive Problem Solving (TRIZ) -based [5] patent classification is essential for patent management and industrial analysis. However, the most research of patent classification are just on English texts and their methods just consider word order but do not use any parse structure, which rely on the word segmentation. In this paper, we propose a new structured representation [11] framework for TRIZ-based Chinese patent classification, which can discover structures automatically and measure structures Our framework uses structured representation model based on reinforcement learning to discover structures. Furthermore, the structured representation model we use lack attention in important structures which may have information memory loss when dealing with long sentences. Therefore, we use attention mechanisms to measure structures that contribute to the patent classification. On 4 TRIZ based classification tasks, our framework significantly outperformed all models in terms of area under curve (AUC) [23] and outperforms Hierarchically Structured LSTM (HS-LSTM) [11]. Moreover, we achieved absolute improvements of 10.08% in performance on "Innovation in product design" classification task in AUC score compared with the state-of-the-art model, bidirectional encoder representation from transformers (BERT) [8].

Automatic fake news detection is a challenging problem which needs a number of verifiable facts support back. Wang et al. [16] introduced LIAR, a validated dataset, and presented a six classes classification task with several popular machine learning methods to detect fake news in linguistic level. However, empirical results have shown that the CNN and RNN based model can not perform very well especially when integrating all features with claim. In this paper, we are the first to present a method to build up a BERT-based [4] mental model to capture the mental feature in fake news detection. In details, we present a method to construct a patterned text in linguistic level to integrate the claim and features appropriately. Then we fine-tune the BERT model with all features integrated text. Empirical results show that our method provides significant improvement over the state-of-art model based on the LIAR dataset we have known by 16.71% in accuracy.

Automatic patent classification based on the TRIZ inventive principles is essential for patent management and industrial analysis. However, acquiring labels for deep learning methods is extraordinarily difficult and costly. This paper proposes a new two-stage semi-supervised learning framework called TRIZ-ESSL, which stands for Enhanced Semi-Supervised Learning for TRIZ. TRIZ-ESSL makes full use of both labeled and unlabeled data to improve the prediction performance. TRIZ-ESSL takes the advantages of semi-supervised sequence learning and mixed objective function, a combination of cross-entropy, entropy minimization, adversarial and virtual adversarial loss functions. Firstly, TRIZ-ESSL uses unlabeled data to train a recurrent language model. Secondly, TRIZ-ESSL initializes the weights of the LSTM-based model with the pre-trained recurrent language model and then trains the text classification model using mixed objective function on both labeled and unlabeled sets. On 3 TRIZ-based classification tasks, TRIZ-ESSL outperforms the current popular semi-supervised training methods and Bert in terms of accuracy score.

The model of attribute information granule based on the method of attribute theory and the problem of fuzzification of attribute information granules are discussed. The concept of information granule given by Zadeh can be explained with an attribute and its qualitative mapping operator. Finally, this study also discusses the reasoning of granulation form of attribute fuzzification information which is based on the fuzzy Petri net, with the good formalisation structure of fuzzy Petri net, as well as its asynchronous, concurrency, uncertainty, and other characteristics which is similar to those of human cognitive activities, enabling the fuzzy Petri net to express the basic characteristics of a cognitive system in the form of computing attribute granules. The results of this study can provide a reference for the establishment of the granular logic model in the uncertain problems and a new interpretation framework for revealing the inherent laws of knowledge uncertainty's change with knowledge granularity, and also provides another new possible approach to the establishment of the machine learning model of fuzzy Petri net.