Search

Time series data recorded from physiological systems often innately exhibit inherent physiological complexity variation on a long-range temporal scale. Multiscale analysis is considered vital for characterising the features of physiological signals. In this research, we propose a novel multiscale analysis method called multiscale increment entropy (MIE), which integrates incremental entropy (IncrEn) and multiscale analysis. MIE inherits the nature of IncrEn, which considers the fluctuation directions and amplitude of a time series. Experiments on both synthetic and real-world signals indicate that MIE performs better than popular approaches as a complexity index. On each temporal scale, MIE corroborates the complexity-loss theory of ageing and disease well. Furthermore, it reliably discriminates either between the EEG time series and heartbeat intervals of healthy subjects and patients or between the oxygen saturation variability of young and elderly, while commonly used algorithms do not perform well in the above cases. MIE requires less computational time compared to several popular approaches. It also has lower variations and is always defined across scales, even for short time series. These merits make it suitable for analysing unknown physiological time series.

Shoeprints contain valuable information for tracing evidence in forensic scenes, and they need to be generated into cleaned, sharp, and high-fidelity images. Most of the acquired shoeprints are found with low quality and/or in distorted forms. The high-fidelity shoeprint generation is of great significance in forensic science. A wide range of deep learning models has been suggested for super-resolution, being either generalized approaches or application specific. Considering the crucial challenges in shoeprint based processing and lacking specific algorithms, we proposed a deep learning based GUV-Net model for high-fidelity shoeprint generation. GUV-Net imitates learning features from VAE, U-Net, and GAN network models with special treatment of absent ground truth shoeprints. GUV-Net encodes efficient probabilistic distributions in the latent space and decodes variants of samples together with passed key features. GUV-Net forwards the learned samples to a refinement-unit proceeded to the generation of high-fidelity output. The refinement-unit receives low-level features from the decoding module at distinct levels. Furthermore, the refinement process is made more efficient by inverse-encoded in high dimensional space through a parallel inverse encoding network. The objective functions at different levels enable the model to efficiently optimize the parameters by mapping a low quality image to a high-fidelity one by maintaining salient features which are important to forensics. Finally, the performance of the proposed model is evaluated against state-of-the-art super-resolution network models.

This paper presents a hardware-friendly compression method for deep neural networks. This method effectively combines layered channel pruning with quantization by a power exponential of 2. While keeping a small decrease in the accuracy of the network model, the computational resources for neural networks to be deployed on the hardware are greatly reduced. These computing resources for hardware resolution include memory, multiple accumulation cells (MACs), and many logic gates for neural networks. Layered channel pruning groups the different layers by decreasing the model accuracy of the pruned network. After pruning each layer in a specific order, the network is retrained. The pruning method in this paper sets a parameter, that can be adjusted to meet different pruning rates in practical applications. The quantization method converts high-precision weights to low-precision weights. The latter are all composed of 0 and powers of 2. In the same way, another parameter is set to control the quantized bit width, which can also be adjusted to meet different quantization precisions. The hardware-friendly pruning quantization (HFPQ) method proposed in this paper trains the network after pruning and then quantizes the weights. The experimental results show that the HFPQ method compresses VGGNet, ResNet and GoogLeNet by 30+ times while reducing the number of FLOPs by more than 85%.

Clustering data with both numeric and categorical attributes is of great importance as such data are ubiquitous in real-world problems. Multi-view learning approaches have proven to be more effective and having better generalisation ability compared to single-view learning in many problems. However, most of the existing clustering algorithms developed for mixed numeric and categorical data are single-view. In this research, we propose a novel multi-view clustering algorithm based on the k-prototypes (which we term Multi-view K-Prototypes) for clustering mixed data. To the best of our knowledge, our proposed Multi-view K-Prototypes is the first multi-view version of the well-known k-prototypes algorithm. To cluster the mixed data over multiple views, we present a novel representation prototype of cluster centres in the scenario of multiple views, and we also devise formulas for updating the cluster centres over each view. Then we propose the concept of consensus cluster centres to output the final clustering result. Finally, we carried out a series of experiments on four benchmark datasets to assess the performance of the proposed Multi-view K-Prototypes clustering. Experimental results show that the Multi-view K-Prototypes algorithm outperforms the seven state-of-the-art algorithms in most cases.

In recent years, hazardous e-mails arose, such as the e-mails infected with ‘viruses’ or ‘worms’ spreading destructive programs and the ‘Phishing Mails’ defrauding e-mail accounts of the users. The number of spams continue to grow. With the related problems of spam coming to be more severe, the spam topics have become significant in various research domains. The common filtering methods include black/white list, rule learning, and those based on text classification, such as Naive Bayes, support vector machine, and boosting trees, multi-agent and genetic algorithm. Among these, the methods based on text classification are most widely applied. Moreover, some efficient methods were proposed to consider only the e-mail's header section, based on which both operating efficiency and classification efficiency could be improved. By applying machine learning technique and decision tree data mining algorithm C4.5, this study aims to propose an efficient spam filtering method with the following features: (i) proposing a two-phase filtering mechanism to scan mainly e-mail's header and auxiliary content. (ii) Reducing the problem of false positive. The experimental results show that the authors’ method has a considerably high accuracy rate of 98.76%. Compared with some other methods of using the same spam data sets or of deep learning-based, their method obviously has an excellent performance.

Clustering is an important unsupervised learning method in machine learning and data mining. Many existing clustering methods may still face the challenge in self-identifying clusters with varying shapes, sizes and densities. To devise a more generic clustering method that considers all the aforementioned properties of the natural clusters, we propose a novel clustering algorithm named Anchor Points based Clustering (APC). The anchor points in APC are characterized by having a relatively large distance from data points with higher densities. We take anchor points as centers to obtain intermediate clusters, which can divide the whole dataset more appropriately so as to better facilitate further grouping. In essence, based on the analysis of the identified anchor points, the relationship among the corresponding intermediate clusters can be better revealed. In short, the difference in local densities (densities within neighboring data points) of the anchor points characterizes their different properties, that is to say, all the intermediate clusters may fall into one or multiple identified levels with different densities. Finally, based on the properties of anchor points, APC spontaneously chooses the appropriate clustering strategies and reports the final clustering results. To evaluate the performances of APC, we conduct experiments on twelve two-dimensional synthetic datasets and twelve multi-dimensional real-world datasets. Moreover, we also apply APC to the Olivetti Face dataset to further assess its effectiveness in terms of face recognition. All experimental results indicate that APC outperforms four classical methods and two state-of-the-art methods in most cases. AI Singapore Ministry of Health (MOH) National Research Foundation (NRF) Accepted version This research is supported by the National Natural Science Foundation of China (61772227,61572227), the Science & Technology Development Founda- tion of Jilin Province (20180201045GX) and the Social Science Foundation of Education Department of Jilin Province (JJKH20181315SK). This research is also supported, in part, by the National Research Foundation Singapore under its AI Singapore Programme (Award Number: AISG-GC-2019-003), the Singapore Ministry of Health under its National Innovation Challenge on Active and Confident Ageing (NIC Project No. MOH/NIC/COG04/2017), and the Joint NTU-WeBank Research Centre on Fintech, Nanyang Technological University, Singapore.

Inferring dynamic system models from observed time course data is very challenging compared to static system identification tasks. Dynamic system models are complicated to infer due to the underlying large search space and high computational cost for simulation and verification. In this research we aim to infer both the structure and parameters of a dynamic system simultaneously by particle swarm optimization (PSO) improved by efficient stratified sampling approaches. More specifically, we enhance PSO with two modern stratified sampling techniques, i.e., Latin hyper cube sampling (LHS) and Latin hyper cube multi dimensional uniformity (LHSMDU). We propose and evaluate two novel swarm-inspired algorithms, LHS-PSO and LHSMDU-PSO, which can be used particularly to learn the model structure and parameters of complex dynamic systems efficiently. The performance of LHS-PSO and LHSMDU-PSO is further compared with the original PSO and genetic algorithm (GA). We chose real-world cancer biological model called Kinetochores to asses the learning performance of LHSMDU-PSO and LHS-PSO in comparison with GA and the original PSO. The experimental results show that LHSMDU-PSO can find promising models with reasonable parameters and structure, and it outperforms LHS-PSO, PSO, and GA in our experiments.

Density peak clustering (DPC) is a recently developed density-based clustering algorithm that achieves competitive performance in a non-iterative manner. DPC is capable of effectively handling clusters with single density peak (single center), i.e., based on DPC’s hypothesis, one and only one data point is chosen as the center of any cluster. However, DPC may fail to identify clusters with multiple density peaks (multi-centers) and may not be able to identify natural clusters whose centers have relatively lower local density. To address these limitations, we propose a novel clustering algorithm based on a hierarchical approach, named multi-center density peak clustering (McDPC). Firstly, based on a widely adopted hypothesis that the potential cluster centers are relatively far away from each other. McDPC obtains centers of the initial micro-clusters (named representative data points) whose minimum distance to the other higher-density data points are relatively larger. Secondly, the representative data points are autonomously categorized into different density levels. Finally, McDPC deals with micro-clusters at each level and if necessary, merges the micro-clusters at a specific level into one cluster to identify multi-center clusters. To evaluate the effectiveness of our proposed McDPC algorithm, we conduct experiments on both synthetic and real-world datasets and benchmark the performance of McDPC against other state-of-the-art clustering algorithms. We also apply McDPC to perform image segmentation and facial recognition to further demonstrate its capability in dealing with real-world applications. The experimental results show that our method achieves promising performance. AI Singapore National Research Foundation (NRF) Accepted version This research is supported by the National Natural Science Foundation of China (61772227,61572227), the Science & Technology Development Foundation of Jilin Province (20180201045GX) and the Social Science Foundation of Education Department of Jilin Province (JJKH20181315SK). This research is also supported, in part, by the National Research Foundation Singapore under its AI Singapore Programme (Award Number: AISG- GC-2019-003), the Singapore Ministry of Health under its National Innovation Challenge on Active and Con dent Ageing (NIC Project No. MOH/NIC/COG04/2017), and the Joint NTU- WeBank Research Centre on Fintech, Nanyang Technological University, Singapore.

Recently, many Deep Learning models have been employed to classify different kinds of plant diseases, but very little work has been done for disease severity detection. However, it is more important to master the severities of plant diseases accurately and timely, as it helps to make effective decisions to protect the plants from being further infected and reduce financial loss. In this paper, based on the Huanglongbing (HLB)-infected leaf images obtained from PlantVillage and crowdAI , we created a dataset with 5,406 citrus leaf images infected by HLB. Then six different kinds of popular models were trained to perform the severity detection of citrus HLB with the goal to find which types of models are more suitable to detect HLB severity with the same training circumstance. The experimental results show that the Inception_v3 model with epochs=60 can achieve higher accuracy than that of other models for severity detection with an accuracy of 74.38% due to its highly computational efficiency and small number of parameters. Additionally, aiming for evaluating whether GANs-based data augmentation can contribute to improve the model learning performance, we adopted DCGANs (Deep Convolutional Generative Adversarial Networks) to augment the original training dataset up to two times itself. Finally, a new training dataset with 14,056 leaf images composed by the original training images and the augmented ones were used to train the Inception_v3 model. As a result, we achieved an accuracy of 92.60%, about 20% higher than that of the Inception_v3 model trained by the original training dataset, which suggested that the GANs-based data augmentation is very useful to improve the model learning performance.

In recent years, convolutional neural network (CNN)-based artificial intelligence algorithms have been widely applied to object recognition and image classification tasks. However, the high performance of convolutional neural networks comes at the cost of high-intensity computing and enormous numbers of parameters, which pose substantial challenges to terminal implementations. An end-to-end FPGA-based accelerator is proposed in this work that efficiently processes fine-grained pruned CNNs. A group pruning algorithm with group sparse regularization (GSR) is introduced to solve internal buffer misalignments and load imbalances of the accelerator after fine-grained pruning. A mathematical model of accelerator access and transmission is established to explore the optimal design scale and calculation mode. The accelerator is optimized by designing sparse processing elements and by scheduling the on- and off-chip buffers. The proposed approach reduces the computation of a state-of-the-art large-scale CNN, VGG16, by 86.9% with an accuracy loss on CIFAR-10 of only 0.48%. The accelerator achieves 188.41 GOPS at 100 MHz and consumes 8.15 W when implemented on a Xilinx VC707, making it more energy-efficient than previous approaches.

Internet of things applications call for increasingly demanding accuracy of indoor positioning. In this work, piezoelectric micromachined ultrasonic transducer (PMUT) based time difference of arrival (TDoA) indoor positioning technology is proposed. Compared with traditional acoustic positioning technology, the technology features small footprint and easy integration into miniaturized systems. We designed and fabricated a low-frequency PMUT chip for demonstration. At the same time, an ultrasonic coding scheme based on binary phase shift keying (BPSK) was used. Finally, a time difference ranging experiment was carried out, and within a distance of 50cm, the error was at the millimeter level. The experimental results verify the feasibility of applying PMUT to indoor positioning systems.

Microrobotics have emerged as a fascinating field with great potential for drug delivery, non-invasive microsurgery, and precise materials assembly. Acoustic devices have been exploited as effective strategies for developing underwater microrobotics. This work presents an integrated wirelessly powered microelectromechanical swimming robot driven by Lamb Wave Resonators (LWRs). Improvements have been made to the structure of LWR driver. The asymmetric distribution of reflectors and leakage hole of the LWR leads to the difference of acoustic wave transmission. Unidirectional drive was realized with the asymmetric acoustic streaming effect. Aided by programming the energy distribution of the LWRs, flexible 2-D motions can be realized, demonstrating an artificial acoustic compartment for tiny swimmers.

The most common issues for many clustering algorithms include the slow convergence, requirement for pre-specification of a number of parameters, and the lack of robustness when dealing with anomalies. Recently, the density peak clustering (DPC) algorithm was proposed to discover the centers of clusters by finding the density peaks in a dataset based on their local densities. The DPC needs neither an iterative process nor a large number of parameters, and it supports a heuristic approach, known as the decision graph, to manually select cluster centroids. However, the selection of the key parameters of the DPC was not systematically investigated. In this paper, we propose the feasible residual error-based density peak clustering algorithm with the fragment merging strategy, where the local density within the neighborhood region is measured through the residual error computation and the resulting residual errors are then used to generate residual fragments for cluster formation. The model parameters are then able to be calculated from the equations with statistical theoretical justification. We also develop a semi-automatic cluster identification method to eliminate the iterative process of manual centroid selection. The robustness and effectiveness of the proposed algorithm compared to the DPC and other clustering algorithms are demonstrated through experiments on standard benchmark datasets. The proposed method named feasible residual error-based density peak clustering (FREDPC) algorithm with the fragment merging strategy only needs to perform in one single step without any iteration and thus it is fast and has a great potential to be applied on a wide range of applications.

Recently, the study of graph neural network (GNN) has attracted much attention and achieved promising performance in molecular property prediction. Most GNNs for molecular property prediction are proposed based on the idea of learning the representations for the nodes by aggregating the information of their neighbor nodes (e.g. atoms). Then, the representations can be passed to subsequent layers to deal with individual downstream tasks. Therefore, the architectures of GNNs can be considered as being composed of two core parts: graph-related layers and task-specific layers. Facing real-world molecular problems, the hyperparameter optimization for those layers are vital. Hyperparameter optimization (HPO) becomes expensive in this situation because evaluating candidate solutions requires massive computational resources to train and validate models. Furthermore, a larger search space often makes the HPO problems more challenging. In this research, we focus on the impact of selecting two types of GNN hyperparameters, those belonging to graph-related layers and those of task-specific layers, on the performance of GNN for molecular property prediction. In our experiments. we employed a state-of-the-art evolutionary algorithm (i.e., CMA-ES) for HPO. The results reveal that optimizing the two types of hyperparameters separately can gain the improvements on GNNs' performance, but optimising both types of hyperparameters simultaneously will lead to predominant improvements. Meanwhile, our study also further confirms the importance of HPO for GNNs in molecular property prediction problems.

In recent years, graph neural networks (GNNs) have gained increasing attention, as they possess the excellent capability of processing graph-related problems. In practice, hyperparameter optimisation (HPO) is critical for GNNs to achieve satisfactory results, but this process is costly because the evaluations of different hyperparameter settings require excessively training many GNNs. Many approaches have been proposed for HPO, which aims to identify promising hyperparameters efficiently. In particular, the genetic algorithm (GA) for HPO has been explored, which treats GNNs as a black-box model, of which only the outputs can be observed given a set of hyperparameters. However, because GNN models are sophisticated and the evaluations of hyperparameters on GNNs are expensive, GA requires advanced techniques to balance the exploration and exploitation of the search and make the optimisation more effective given limited computational resources. Therefore, we proposed a tree-structured mutation strategy for GA to alleviate this issue. Meanwhile, we reviewed the recent HPO works, which gives room for the idea of tree-structure to develop, and we hope our approach can further improve these HPO methods in the future.

Explainable Artificial Intelligence (XAI) is rapidly becoming an emerging and fast-growing research field; however, its adoption in healthcare is still at the early stage despite the potential that XAI can bring to the application of AI in this industry. Many challenges remain to be solved, including setting standards for explanations, the degree of interaction between different stakeholders and the models, the implementation of quality and performance metrics, the agreement on standards for safety and accountability, its integration into clinical workflows, and IT infrastructure. This paper has two objectives. The first one is to present summarized outcomes of a literature survey and highlight the state-of-the-art for explainability including gaps, challenges, and opportunities for XAI in healthcare industry. For easier comprehension and onboarding to this research field we suggest a synthesized taxonomy for categorizing explainability methods. The second objective is to ask the question if applying a novel way of looking at explainability problem space, through a specific problem/domain lens, and automating that approach in an AutoML similar fashion, would help mitigate the challenges mentioned above. In the literature there is a tendency to look at the explainability of AI from model-first lens, which puts concrete problems and domains aside. For example, the explainability of a patient's survival model is treated the same as explaining a hospital cost procedure calculation. With a well-identified problem/domain that XAI should be applied to, the scope is clear and well-defined, enabling us to (semi-) automatically find suitable models, optimize their parameters and their explanations, metrics, stakeholders, safety/accountability level, and suggest means of their integration into clinical workflow .

Mild cognitive impairment (MCI) has been described as the intermediary stage before Alzheimer's Disease - many people however remain stable or even demonstrate improvement in cognition. Early detection of progressive MCI (pMCI) therefore can be utilised in identifying at-risk individuals and directing additional medical treatment in order to revert conversion to AD as well as provide psychosocial support for the person and their family. This paper presents a novel solution in the early detection of pMCI people and classification of AD risk within MCI people. We proposed a model, MudNet, to utilise deep learning in the simultaneous prediction of progressive/stable MCI classes and time-to-AD conversion where high-risk pMCI people see conversion to AD within 24 months and low-risk people greater than 24 months. MudNet is trained and validated using baseline clinical and volumetric MRI data (n = 559 scans) from participants of the Alzheimer's Disease Neuroimaging Initiative (ADNI). The model utilises T1-weighted structural MRIs alongside clinical data which also contains neuropsychological (RAVLT, ADAS-11, ADAS-13, ADASQ4, MMSE) tests as inputs. The averaged results of our model indicate a binary accuracy of 69.8% for conversion predictions and a categorical accuracy of 66.9% for risk classifications.

Convolutional neural networks (CNNs) have achieved a wide range of applications in the field of computer vision. However, its large number of parameters and computation limits its deployment in embedded terminals. Designing dedicated CNN accelerators becomes an efficient way to solve the situation. In this paper, a basic access model of systolic accelerator is established to find the optimal design parameters for specific CNN structure. The experiment result shows that $14 \times16$ is the optimum size of processing element array for AlexNet and VGG16, which matches with the design parameters of accelerators with the same architecture.

With the extensive application of information acquisition and transmission technology in power system, the extraction of equipment basic data becomes feasible. From the perspective of lean operation and maintenance, this paper systematically combs the scale structure, equipment quality dimension, equipment efficiency dimension and operation and maintenance cost dimension of distribution network equipment assets, and uses correlation analysis and principal component analysis to deeply mine the index data to study the power grid Data relationship of equipment related information. A data stream cleaning method based on sliding window is proposed to detect abnormal data sequence and clean data. The clustering factors are calculated by principal component analysis method, and 55 kinds of data are reduced. Finally, a set of 16 principal component data processing flow of distribution network equipment status is constructed. Finally, the feasibility and effectiveness of the evaluation index system and evaluation method are verified by comprehensive evaluation of the equipment status of a certain distribution network.

With the development of convolutional neural network (CNN), the accuracy of face recognition has been greatly improved. But the huge amount of weights and calculations hinders its implementation in portable devices. Designing hardware accelerator is an effective solution to the problem. In this paper, a face recognition algorithm is designed based on deep separable convolution. The weights and activations are quantified to 8 bits, reducing the requirement of data access and bandwidth. In addition, a generic CNN accelerator based on systolic array is designed and validated on Xilinx Zynq-XC7Z035 FPGA. The face recognition algorithm achieved an accuracy of 94.4% in the LFW dataset. The performance and power efficiency of the accelerator are 52.9 GOPS and 9.71GOPS/W at 100MHz, respectively. And the accelerator can process 160×160 face image at 25FPS.

To better address the recognition of abnormalities among mammographic images, in this study we apply the deep fusion learning approach based on Pre-trained models to discover the discriminative patterns between Normal and Tumor categories. We designed a deep fusion learning framework for mammographic image classification. This framework works in two main steps. After obtaining the regions of interest (ROIs) from original dataset, the first step is to train our proposed deep fusion models on those ROI patches which are randomly collected from all ROIs. We proposed the deep fusion model (Model1) to directly fuse the deep features to classify the Normal and Tumor ROI patches. To explore the association among channels of the same block, we propose another deep fusion model (Model2) to integrate the cross-channel deep features using 1 × 1 convolution. The second step is to obtain the final prediction by performing the majority voting on all patches' prediction of one ROI. The experimental results show that Model1 achieves the whole accuracy of 0.8906, recall rate of 0.913, and precision rate of 0.8077 for Tumor class. Accordingly, Model2 achieves the whole accuracy of 0.875, recall rate of 0.9565, and precision rate 0.7,586 for Tumor class. Finally, we open source our Python code at https://github.com/yxchspring/MIAS in order to share our tool with the research community.

In fields such as finance, medicine, engineering, and science, making real-time predictions during transient periods characterized by sudden and large changes is a hard challenge for machine learning. Humans keep memory of these transient events, abstractly learn the most relevant rules and reuse them when similar events occur, which stems from episodic memory that allows storage and recall of similar events. This paper proposes a novel online general episodic memory mechanism (GEMM) and demonstrates its integration into the Neuro-Fuzzy system (NFS) architecture called evolving Mamdani Fuzzy Inference System (eMFIS) with Fuzzy Rule Interpolation and Extrapolation (FRI/E). Our proposition, called GEMM-eMFIS(FRI/E), learns from past events by storing and retrieving them from an episodic memory cache during event-driven transient behavior, thereby boosting performance while using a few rules only. GEMM-eMFIS(FRI/E) further has several in-built mechanisms that enable it to learn effectively from continuous stream of online data. They include associative-dissociative learning theory to keep its rule base updated, 2-stage incremental clustering; (2SIC) to determine cluster width, interpolation and extrapolation of rules to deal with concept shifts and drifts in the time-variant data, and rule pruning and merging to keep the rule base compact. GEMM-eMFIS (FRI/E) is benchmarked against other NFS’ on various time-variant datasets such as stock index prices and rainfall runoff with 3%-5% improvement during transient period and shows strong forecasting performances with 4%-5% more interpretability with lesser rules.

Numerous human actions such as “Phoning,” “PlayingGuitar,” and “RidingHorse” can be inferred by static cue-based approaches even if their motions in video are available considering one single still image may already sufficiently explain a particular action. In this research, we investigate human action recognition in still images and utilize deep ensemble learning to automatically decompose the body pose and perceive its background information. Firstly, we construct an end-to-end NCNN-based model by attaching the nonsequential convolutional neural network (NCNN) module to the top of the pretrained model. The nonsequential network topology of NCNN can separately learn the spatial- and channel-wise features with parallel branches, which helps improve the model performance. Subsequently, in order to further exploit the advantage of the nonsequential topology, we propose an end-to-end deep ensemble learning based on the weight optimization (DELWO) model. It contributes to fusing the deep information derived from multiple models automatically from the data. Finally, we design the deep ensemble learning based on voting strategy (DELVS) model to pool together multiple deep models with weighted coefficients to obtain a better prediction. More importantly, the model complexity can be reduced by lessening the number of trainable parameters, thereby effectively mitigating overfitting issues of the model in small datasets to some extent. We conduct experiments in Li’s action dataset, uncropped and 1.5x cropped Willow action datasets, and the results have validated the effectiveness and robustness of our proposed models in terms of mitigating overfitting issues in small datasets. Finally, we open source our code for the model in GitHub (https://github.com/yxchspring/deep_ensemble_learning) in order to share our model with the community.

The Guesser is a task of visual grounding in GuessWhat?! like visual dialogue. It locates the target object in an image supposed by an Oracle oneself over a question-answer based dialogue between a Questioner and the Oracle. Most existing guessers make one and only one guess after receiving all question-answer pairs in a dialogue with the predefined number of rounds. This paper proposes a guessing state for the Guesser, and regards guess as a process with change of guessing state through a dialogue. A guessing state tracking based guess model is therefore proposed. The guessing state is defined as a distribution on objects in the image. With that in hand, two loss functions are defined as supervisions to guide the guessing state in model training. Early supervision brings supervision to Guesser at early rounds, and incremental supervision brings monotonicity to the guessing state. Experimental results on GuessWhat?! dataset show that our model significantly outperforms previous models, achieves new state-of-the-art, especially the success rate of guessing 83.3% is approaching the human-level accuracy of 84.4%.

Objectives: The purpose of this research is to explore innovative interaction mode of children’s marker education, to solve the problems of universal homogeneity of Marker education products and lack of emotional communication, and to develop children’s intelligent interactive toys by using open source hardware coding technology. Methods: Focusing on the three elements of user-interaction-technology, this study comprehensively analyzed the needs and problems of educational play aids for markers by means of questionnaires, behavioral observation and competitive analysis, and carried out product innovation design through design practice. Results: The research ultimately establishes the brand “KOBOT” multi-mode interactive experience marker education toy, which is a set of innovative marker education products based on the integration of various offline entity modules and online programming applications. It supports a variety of modeling modular design, interactive technology and interactive ways. Conclusions: “KOBOT” marker education product expands the game process of construction mode and intellectuality, enhances the interaction between children and marker products, fosters the social awareness of parent-child interaction and friend cooperation, and lays the development direction for the future of Chinese children’s marker education.

As the important elements of the Internet of Things system, wireless sensor network (WSN) has gradually become popular in many application fields. However, due to the openness of WSN, attackers can easily eavesdrop, intercept, and rebroadcast data packets. WSN has also faced many other security issues. Intrusion detection system (IDS) plays a pivotal part in data security protection of WSN. It can identify malicious activities that attempt to violate network security goals. Therefore, the development of effective intrusion detection technologies is very important. However, many dimensions of the datasets of IDS are irrelevant or redundant. This causes low detection speed and poor performance. Feature selection is thus introduced to reduce dimensions in IDS. At the same time, many evolutionary computing (EC) techniques were employed in feature selection. However, these techniques usually have just one Candidate Solution Generation Strategy (CSGS) and often fall into local optima when dealing with feature selection problems. The self-adaptive differential evolution (SaDE) algorithm is adopted in our paper to deal with feature selection problems for IDS. The adaptive mechanism and four effective CSGSs are used in SaDE. Through this method, an appropriate CSGS can be selected adaptively to generate new individuals during evolutionary process. Besides, we have also improved the control parameters of the SaDE. The K-Nearest Neighbour (KNN) is used for performance assessment for feature selection. KDDCUP99 dataset is employed in the experiments, and experimental results demonstrate that SaDE is more promising than the algorithms it compares.

This study investigates the feasibility of applying state of the art deep learning techniques to detect precancerous stages of squamous cell carcinoma (SCC) cancer in real time to address the challenges while diagnosing SCC with subtle appearance changes as well as video processing speed. Two deep learning models are implemented, which are to determine artefact of video frames and to detect, segment and classify those no-artefact frames respectively. For detection of SCC, both mask-RCNN and YOLOv3 architectures are implemented. In addition, in order to ascertain one bounding box being detected for one region of interest instead of multiple duplicated boxes, a faster non-maxima suppression technique (NMS) is applied on top of predictions. As a result, this developed system can process videos at 16-20 frames per second. Three classes are classified, which are ‘suspicious’, ‘high grade’ and ‘cancer’ of SCC. With the resolution of 1920x1080 pixels of videos, the average processing time while apply YOLOv3 is in the range of 0.064-0.101 seconds per frame, i.e. 10-15 frames per second, while running under Windows 10 operating system with 1 GPU (GeForce GTX 1060). The averaged accuracies for classification and detection are 85% and 74% respectively. Since YOLOv3 only provides bounding boxes, to delineate lesioned regions, mask-RCNN is also evaluated. While better detection result is achieved with 77% accuracy, the classification accuracy is similar to that by YOLOYv3 with 84%. However, the processing speed is more than 10 times slower with an average of 1.2 second per frame due to creation of masks. The accuracy of segmentation by mask-RCNN is 63%. These results are based on the date sets of 350 images. Further improvement is hence in need in the future by collecting, annotating or augmenting more datasets.

Swarm intelligence algorithms (SIAs) have demonstrated excellent performance when solving optimization problems including many real-world problems. However, because of their expensive computational cost for some complex problems, SIAs need to be accelerated effectively for better performance. This paper presents a high-performance general framework to accelerate SIAs (FASI). Different from the previous work which accelerates SIAs through enhancing the parallelization only, FASI considers both the memory architectures of hardware platforms and the dataflow of SIAs, and it reschedules the framework of SIAs as a converged dataflow to improve the memory access efficiency. FASI achieves higher acceleration ability by matching the algorithm framework to the hardware architectures. We also design deep optimized structures of the parallelization and convergence of FASI based on the characteristics of specific hardware platforms. We take the quantum behaved particle swarm optimization algorithm as a case to evaluate FASI. The results show that FASI improves the throughput of SIAs and provides better performance through optimizing the hardware implementations. In our experiments, FASI achieves a maximum of 290.7 Mb/s throughput which is higher than several existing systems, and FASI on FPGAs achieves a better speedup than that on GPUs and multi-core CPUs. FASI is up to 123 times and not less than 1.45 times faster in terms of optimization time on Xilinx Kintex Ultrascale xcku040 when compares to Intel Core i7-6700 CPU/NVIDIA GTX1080 GPU. Finally, we compare the differences of deploying FASI on hardware platforms and provide some guidelines for promoting the acceleration performance according to the hardware architectures.

Fireworks algorithm (FWA) is a novel swarm intelligence algorithm recently proposed for solving complex optimization problems. Because of its powerful global optimization ability to solve classification problems, we first present an optimization classification model in this paper. In this model, a linear equation set is constructed according to classification problems. This optimization classification model can be solved by most evolutionary computation techniques. In this paper, a self-adaptive FWA (SaFWA) is developed so that the optimization classification model can be solved efficiently. In SaFWA, four candidate solution generation strategies (CSGSs) are employed to increase the diversity of solutions. In addition, a self-adaptive search mechanism has also been introduced to use the four CSGSs simultaneously. To extensively assess the performance of SaFWA on solving classification problems, eight datasets have been used in the experiments. The experimental results show that it is feasible to solve classification problems through the optimization classification model and SaFWA. Furthermore, SaFWA performs better than FWA, FWA variants with only one CSGS, particle swarm optimization, and differential evolution on most of the training sets and test sets.

Support vector clustering (SVC) is a boundary-based algorithm, which has several advantages over other clustering methods, including identifying clusters of arbitrary shapes and numbers. Leveraged by the high generalization ability of the large margin distribution machine (LDM) and the optimal margin distribution clustering (ODMC), we propose a new clustering method: minimum distribution for support vector clustering (MDSVC), for improving the robustness of boundary point recognition, which characterizes the optimal hypersphere by the first-order and second-order statistics and tries to minimize the mean and variance simultaneously. In addition, we further prove, theoretically, that our algorithm can obtain better generalization performance. Some instructive insights for adjusting the number of support vector points are gained. For the optimization problem of MDSVC, we propose a double coordinate descent algorithm for small and medium samples. The experimental results on both artificial and real datasets indicate that our MDSVC has a significant improvement in generalization performance compared to SVC.

The speech keyword detection system is generally used to detect a specific wake-up word spoken by the user, as a simple interactive command between the user and the device, or a front end of a larger voice recognition system for determining whether to open a large speech recognition system at the back end. Yet, such keyword recognition systems are always active, hence have strict power and latency constraints. This paper proposes an implementation of an LSTM accelerator based on an iterative linear approximation scheme, which optimizes the circuit implementation while maintaining high accuracy and reliability. The scheme of this paper has been verified on the FPGA board, and real-time speech recognition power consumption is only less 150$\mu$w.

In this paper we propose DeepSwarm, a novel neural architecture search (NAS) method based on Swarm Intelligence principles. At its core DeepSwarm uses Ant Colony Optimization (ACO) to generate ant population which uses the pheromone information to collectively search for the best neural architecture. Furthermore, by using local and global pheromone update rules our method ensures the balance between exploitation and exploration. On top of this, to make our method more efficient we combine progressive neural architecture search with weight reusability. Furthermore, due to the nature of ACO our method can incorporate heuristic information which can further speed up the search process. After systematic and extensive evaluation, we discover that on three different datasets (MNIST, Fashion-MNIST, and CIFAR-10) when compared to existing systems our proposed method demonstrates competitive performance. Finally, we open source DeepSwarm (https://github.com/Pattio/DeepSwarm) as a NAS library and hope it can be used by more deep learning researchers and practitioners.

Data clustering is a well-researched area in data mining and machine learning. The clustering algorithms that can handle both numeric and categorical variables have been extensively researched in the recent years. However, the clustering algorithms have a major limitation that converge to a local optima. Therefore, to address this problem this paper has proposed a novel algorithm ABC k-prototypes (Artificial Bee Colony clustering based on k-prototypes) for clustering mixed data. In our proposed approach we use the combination between the distribution centroid and the mean to calculate the dissimilarity between data objects and prototypes. The proposed algorithm is tested on five different datasets taken from the UCI machine learning data repository. The comparative results in the performance measures of the clustering showed that the proposed algorithm outperformed the traditional k-prototypes.

Inferring models of dynamic systems from their time series data is a challenging task for optimization algorithms due to its potentially expensive computational cost and underlying large search space. In this study we aim to infer both the structure and parameters of a dynamic system model simultaneously by Particle Swarm Optimization (PSO), enhanced by effective stratified sampling strategies. More specifically we apply Latin Hyper Cube Sampling (LHS) with PSO. This leads to two novel swarm-inspired algorithms, LHS-PSO which can be used efficiently to learn the structure and parameters of simple and complex dynamic system models. We used a complex biological cancer model called Kinetochores, for assessing the performance of PSO and LHS-PSO. The experimental results demonstrate that LHS-PSO can find promising solutions with corresponding structure and parameters, and it outperforms PSO during our experiments.

Many real-world problems can be naturally formulated as discrete multi-objective optimization (DMOO) problems. In this research we propose a novel bio-inspired Physarum competition algorithm (PCA) to tackle DMOO problems by modelling the Physarum discrete motility over a hexagonal cellular automaton. Our algorithm is based on the chemo-attraction forces towards food resources (Objective Functions) and the repulsion negative forces between the competing Physarum. Numerical experimental work clearly demonstrated that our PCA algorithm had the best performance for the spread indicator against three state-of-the-art evolutionary algorithms, and its effectiveness in terms of commonly used metrics. These results have indicated the superiority of PCA in exploring the search space and keeping diversity, this makes PCA a promising algorithm for solving DMOO problems.

Network pruning is important for the deployment of deep neural network on hardware platforms. However, most pruning methods focus on coarse-grained pruning with high precision loss. In addition, many fine-grained pruning methods focus the pruning on the fully connected layer rather than the convolution layer. The group pruning technique is proposed, focusing only on CONV-layer that enables to keep the weight reduction rate consistent within each weight group. This helps solve some inefficiency problems including internal buffer misalignment and load imbalance after fine-grained pruning. In the pre-training, the strategy of group sparse regularization (GSR) and standardizing the weight distribution is added to alleviate the loss of precision under high sparsity. Finally, we used the MNIST and CIFAR-10 data sets to test the LeNet and VGG-16 for experiments. The reduction rates of convolution layer are 87.5% and 62.5% respectively in the range of 0.14% accuracy loss.

As a challenging computer vision task, fine-grained visual categorization (FGVC) has received more and more attention. Most fine-grained classification algorithms are data-hungry and time-consuming, on the contrary, humans have the ability to recognize different species with only a small number of samples. For example, a baby can identity different kinds of dogs with only few images. Therefore, in this paper, a fined-grained classification model which is based on meta learning is proposed, and it can be used for more accurate fined-grained classification with only a small number of data. Meta learning uses meta-learners to learn relative tasks so that they can handle with new tasks faster and better when facing new tasks. Specifically, our model uses a task learner to learn various tasks and update the parameters of meta learner at the same time. When facing a new task, meta learner can learn and adapt to new tasks more quickly. The key novelty of our model is an attention module which can enlarge the local area response of image. This model has achieved great results on both CUB-200-2011 and Stanford dogs, and the accuracy of this model has achieved 81.74% with only one image of CUB-200-2011.

Convolutional neural network (CNN) has achieved excellent achievements in the field of image recognition. Because of the better energy efficiency and fewer development cycle, FPGA has been widely used as the implementation platform of CNN. The quantized CNN reduces the demand for storage and transmission bandwidth, and is easier to be implemented on FPGA platforms. And the weights of the quantized network contain a lot of repetitions and zeros. The repeatability and sparsity of these weight parameters can be well utilized in the hardware implementation of CNN. In this paper, a CNN accelerator is designed and implemented on the Virtex-7 FPGA, and the test network is AlexNet with 17 unique weights after quantization. By using weight combination method, the accelerator redesigned the calculation pattern of the processing elements (PEs), reducing the multiplication by 92.25%. The experiment result shows that our accelerator achieves 179 GOPS under 200 MHz clock frequency. The energy efficiency of 32.72 GOPS/Watt is obtained by combining the new PE with the systolic array architecture.

In recent years, convolutional neural networks (CNN) has achieved great success in computer vision tasks, such as object detection, face recognition and image classification. With the diversification of CNN models, accelerators that only support a single network can no longer meet the needs of applications. Due to the computational intensiveness of convolution operations, the implementation of CNN on the FPGA platform faces many challenges. In this paper, a convolution unit based on systolic arrays is proposed in the design of CNN accelerator, and the fixed-point quantization method is adopted to save a large amount of storage resources and reduce the required transmission bandwidth, thus improving throughput and power efficiency. The performance density and power efficiency of our design can reach 0.165 GOPS/DSP and 36.3 GOPS/W under 100MHz clock frequency.

Deep Convolutional Neural Networks (DCNN) significantly improve the performance of many computer vision tasks, such as classification, detection, and semantic segmentation. The ideal face features are expected to have smaller maximal intra-class distance than minimal inter-class distance under open-set protocol, but the current algorithm still has the open problem of implementing the criterion. In this paper, we present a feature augmentation network for the IARPA Janus Benchmark C (IJB-C) on a small CNN. The proposed feature enhancement method is used to approximate the identity features, and the original features are augmented by a small automatic encoder-decoder which can be quickly run in an embedded system with limited resources and obtains similar accuracy to a large backbone CNN network.

The operation of a wastewater treatment plant (WWTP) is a typical complex control problem, with nonlinear dynamics and coupling effects among the variables, which renders the implementation of real-time optimal control an enormous challenge. In this study, a Q-learning algorithm with activated sludge model No. 2d-guided (ASM2d-guided) reward setting (an integrated ASM2d-QL algorithm) is proposed, and the widely applied anaerobic-anoxic-oxic (AAO) system is chosen as the research paradigm. The integrated ASM2d-QL algorithms equipped with a self-learning mechanism are derived for optimizing the control strategies (hydraulic retention time (HRT) and internal recycling ratio (IRR)) of the AAO system. To optimize the control strategies of the AAO system under varying influent loads, Q matrixes were built for both HRTs and IRR optimization through the pair of &lt, max reward-action&gt, based on the integrated ASM2d-QL algorithm. 8 days of actual influent qualities of a certain municipal AAO wastewater treatment plant in June were arbitrarily chosen as the influent concentrations for model verification. Good agreement between the values of the model simulations and experimental results indicated that this proposed integrated ASM2d-QL algorithm performed properly and successfully realized intelligent modeling and stable optimal control strategies under fluctuating influent loads during wastewater treatment.

With the progress of the urbanization, a series of traffic problems have occurred because of the growing urban population and the far lower growth rate of roads than that of cars. One of the most prominent problems is traffic congestion problem. The prediction of traffic congestion is the key to alleviate traffic congestion. To ensure the real-time performance and accuracy of the traffic congestion prediction, we propose a short-term traffic congestion prediction model called TCPModel based on deep learning. By processing a massive amount of urban taxi transportation data, we extract the traffic volume and average speed of taxis which are the most important parameters for assessing of traffic flow prediction. After analyzing the temporal and spatial distribution characteristics of the traffic flow and average speed, we present a short-term traffic volume prediction model called TVPModel, and a short-term traffic speed prediction model called TSPModel. Both models are based on a deep learning method Stacked Auto Encoder (SAE). By comparing the other traffic flow forecasting methods and average speed forecasting methods, the methods proposed by this paper have improved the accuracy rate. For traffic congestion recognition, we use a novel model called TCPModel based on three traffic parameters (average speed, traffic flow and density), which uses standard function method to standardize the parameters and calculate the congestion comprehensive threshold to determine the congestion level by thresholds. According to the experiments, TVPModel and TSPModel in this paper got satisfied accuracy compared with other prediction models.

Insect species recognition is an important application of computer vision in zoology and agriculture. Most of existing methods resort to hand-crafted features and traditional classifiers, which usually give poor accuracy and apply only to elaborately taken full-size pictures. In this paper, we focus on a more challenging case where the images are taken in the wild with complex backgrounds, and propose to use a deep learning based detection model to deal with it. It exploits multi-class object detection to eliminate interferences from complex backgrounds, while taking advantages of deep learning to significantly improve the performance of recognition. After evaluating several popular detection methods, R-FCN is selected as the base model. To further improve its performance, we introduce a clustering algorithm for estimation of the anchor boxes instead of using predefined ones. The experimental results on a dataset of insect images collected in the wild prove the effectiveness of our proposed method in improving both accuracy and speed.

With the development of curtain wall technology, the unit curtain wall has been widely used as a decorative component of high-rise buildings, but it has brought great difficulties to the later safety inspection. Based on the theory of thin-slab vibration, the laser vibration measurement technology in this paper is used to analyze the relationship between the multi-order natural frequency and the safety status of the point-supported experimental curtain wall model. At the same time, a corresponding finite element model is established to further quantify the natural frequency and restraining force on the curtain wall. The result shows that there are different relationships between natural frequencies of different orders and the safety of the curtain wall. In actual engineering, it is more sensitive for the first-order natural frequency to changes in the safety status of the curtain wall. A new method of safety detection in high-rise building curtain walls can be achieved by the combination of remote laser vibration measurement technology.

In this paper, we study scheduling problems that have batching considerations, and with sequence-dependent setup times for single machine and job shop flow planning. An important focus is the explicit treatment of setup times (costs), missed due dates (tardiness) and wastage as important cost components that impacts the direct and indirect cost. Indirect cost reflects the true cost of an urgent order that caused the machine setup to be re-arranged that can influence business decision on how to priced such order and how on a recurring basis might affect the bottom line profit and lost.[Allahverdi et al., 2006] found that the majority of the earlier papers assumed that the setup time (cost) is negligible or part of the job processing time (cost). This assumption simplifies the analysis however it adversely affects the solution quality. However, there is an increased interest in scheduling problems involving setup times as many recognised that there are tremendous savings when setup times/costs are explicitly incorporated in scheduling decision in various real world industrial environments.

Feature selection has been widely used in classification for improving classification accuracy and reducing computational complexity. Recently, evolutionary computation (EC) has become an important approach for solving feature selection problems. However, firstly, as the datasets processed by classifiers become increasingly large and complex, more and more irrelevant and redundant features may exist and there may be more local optima in the large-scale feature space. Therefore, traditional EC algorithms which have only one candidate solution generation strategy (CSGS) with fixed parameter values may not perform well in searching for the optimal feature subsets for large-scale feature selection problems. Secondly, many existing studies usually use only one classifier to evaluate feature subsets. To show the effectiveness of evolutionary algorithms for feature selection problems, more classifiers should be tested. Thus, in order to efficiently solve large-scale feature selection problems and to show whether the EC-based feature selection method is efficient for more classifiers, a self-adaptive parameter and strategy based particle swarm optimization (SPS-PSO) algorithm is proposed in this paper using multiple classifiers. In SPS-PSO, a representation scheme of solutions and five CSGSs have been used. To automatically adjust the CSGSs and their parameter values during the evolutionary process, a strategy self-adaptive mechanism and a parameter self-adaptive mechanism are employed in the framework of particle swarm optimization (PSO). By using the self-adaptive mechanisms, the SPS-PSO can adjust both CSGSs and their parameter values when solving different large-scale feature selection problems. Therefore, SPS-PSO has good global and local search ability when dealing with these large-scale problems. Moreover, four classifiers, i.e., k-nearest neighbor (KNN), linear discriminant analysis (LDA), extreme learning machine (ELM), and support vector machine (SVM), are individually used as the evaluation functions for testing the effectiveness of feature subsets generated by SPS-PSO. Nine datasets from the UCI Machine Learning Repository and Causality Workbench are used in the experiments. All the nine datasets have more than 600 dimensions, and two of them have more than 5,000 dimensions. The experimental results show that the strategy and parameter self-adaptive mechanisms can improve the performance of the evolutionary algorithms, and that SPS-PSO can achieve higher classification accuracy and obtain more concise solutions than those of the other algorithms on the large-scale feature problems selected in this research. In addition, feature selection can improve the classification accuracy and reduce computational time for various classifiers. Furthermore, KNN is a better surrogate model compared with the other classifiers used in these experiments.

Simultaneously inferring both the structure and parameters of Ordinary Differential Equations (ODEs) for a complex dynamic system is more practical in many systems identification problems, but it remains challenging due to the complexity of the underlying search space. In this research, we propose a novel algorithm based on Particle Swarm Optimization (PSO) and Latin Hypercube Sampling (LHS) to address the above problem. The proposed algorithm is termed LatinPSO, and it can be effectively used for inferring the structure and parameters of ODE models through time course data. To start with, the real Human Immunodeficiency Virus (HIV) model and several synthetic models are used for evaluating the performance of LatinPSO. Experimental results demonstrated that LatinPSO could find satisfactory candidate ODE models with appropriate structure and parameters.