Search

Knowledge tracing (KT) is the core task of computer-aided education systems, and it aims at predicting whether a student can answer the next exercise (i.e., question) correctly based on his/her historical answer records. In recent years, deep neural network-based approaches have been widely developed in KT and achieved promising results. More recently, several researches further boost these KT models via exploiting plentiful relationships including exercise-skill relations (E-S), the exercise similarity (E-E) as well as skill similarity (S-S). However, these relationship information are frequently absent in many real-world educational applications, and it is a labor-intensive work for human experts to label it. Inspired by recent advances in natural language processing domain, we propose a novel pre-training approach, namely as SPAKT, and utilize self-supervised learning to pre-train exercise embedding representation without the need for expensive human-expert annotations in this paper. Contrary to existing pre-training methods that highly rely on manually labeling knowledge about the E-E, S-S, or E-S relationships, the core idea of the proposed SPAKT is to design three self-attention modules to model the E-S, E-E, and S-S relationships, respectively, and all of these three modules can be trained in the self-supervised setting. As a pre-training approach, our SPAKT can be effortlessly incorporated into existing deep neural network-based KT frameworks. We experimentally show that, even without using expensive annotations about the aforementioned three kinds of relationships, our model achieves competitive performance compared with state-of-the-arts. Our algorithm implementations have been made publicly available at https://github.com/Vinci-hp/pretrainKT.

This paper studies the issue of the adaptive neural security controller design for uncertain networked singular systems in the presence of deception attacks. Considering that the attack signal is unknown, the neural networks technique is exploited to approximate the attack signal, which eliminates the assumption that the attack signal has a known upper bound. By combining the state feedback with the estimated information of the attack, the impact of the attack is effectively compensated. Furthermore, a novel Lyapunov function, including the decomposed state vector and the weight matrix estimation error, is established to evaluate the bounded area of the system state. Finally, a numerical example substantiates the validity of the theoretical results.

Multiple-choice reading comprehension (MCRC) aims to build an intelligent system that automatically selects an answer from a candidate set when given a passage and a question. Existing MCRC systems rarely consider incorporating external knowledge such as explicit semantic information. In this work, we propose a Contextual and Semantic Fusion Network (CSFN) which effectively integrates contextual and semantic representation. CSFN introduces explicit structured semantics from pre-trained semantic role labeling. Specially, we regard explicit semantic representation as an important feature to fuse with contextual representation, which enriches the representation of sentences. By combining with the transfer learning strategy, the CSFN model has better generalization over limited datasets. To evaluate the ability of our model, we conduct experiments on three MCRC benchmark datasets: RACE, DREAM, and MCTest. Experimental results demonstrate the effectiveness of our proposed model.

We propose a new variant of the Benes network using a merge-replace-fold approach. It is to realize a large port-count optical switch with low waveguide crossings. A bidirectional switch is built instead of a unidirectional switch. Compared to the classical Benes, it requires the same number of switches but far fewer intersections. The novel designs enable a 24-percentage reduction in waveguide intersections. Moreover, it has a worst-case insertion loss of 12.01 dB and a comparable crosstalk level of −13.67 dB between the neighbouring paths. Thus, it offers an energy-efficient way to solve the optical on-chip design problem in a bidirectional manner.

Automatic lymphoma detection and accurate lymphoma boundary delineation from whole body Positron Emission Tomography/Computed Tomography (PET/CT) scans are essential for surgical navigation and radiation therapy. Besides, labeling the data, which means contouring the lymphoma contour in images is time-consuming, operator intensive and subjective. Hence, this paper integrates the supervised learning and unsupervised learning to propose an end-to-end segmentation network, namely DenseX-Net, for both lymphoma detection and segmentation. There are two important flows in the proposed DenseX-Net. One is a reconstruction flow (based on the convolutional encoder-decoder form) that can be used for learning semantic representations of different lymphomas by minimizing the discrepancy between each input and its output in an unsupervised learning form. The other one is a segmentation flow (based on DenseU-Net) that performs the lymphoma segmentation task. Note that, the encoders in both flows are trained jointly with the same weights, which can facilitate DenseX-Net obtaining the accurate segmentation using a little labeled data. We evaluate our proposed DenseX-Net for lymphoma segmentation on 80 real PET/CT cases (from General Hospital of Northern Military Area) with a Dice coefficient of 72.84%. Experimentations and comparisons demonstrate the accuracy and robustness of DenseX-Net as well as its performance advantages as compared with related segmentation networks.

High-resolution CT images can clearly display anatomical structures but does not display functional information, while blurred PET images can display molecular and functional information of lesions but cannot clearly display morphological structures. Therefore, accurate PET-CT image registration, which is used for anatomical structure and functional information fusion, is a prerequisite for early stage cancer diagnosis. However, some hypermetabolic anatomical structures, such as brain and bladder, have low registration accuracy. To solve this problem, a 3D unsupervised network based on a metabolic constraint function and a multi-domain similarity measure (3D MC-MDS Net) is proposed for 3D PET-CT image registration. Specifically, a metabolic constraint model is established based on the standard uptake value (SUV) distribution of hypermetabolic regions such as brain, bladder, liver and heart, which reduces the excessive distortion on displacement vector field (DVF) caused by hypermetabolic anatomical structures in PET images. A DVF estimator is built based on 3D unsupervised convolutional neural networks and a spatial transformer is used for warping 3D PET images to 3D CT images. The generated registration results (PET image patches) and the original 3D CT image patches are used for calculating the spatial domain similarity (SD similarity) and frequency domain similarity (FD similarity). Finally, the loss function of the entire registration network is constructed by a weighted sum of SD similarity, FD similarity and a smoothness of DVF. A dataset consisted of 170 whole-body PET-CT images is used for registration accuracy evaluation. The proposed unsupervised registration network, 3D MC-MDS Net, can accurately learn the 3D registration model by using the training dataset with the metabolic constraint model, which significantly improves the registration accuracy.

Algal bloom is a nonlinear and time-varying process, which brings challenges for the accurate prediction. For the existing mechanism model of algae ignores the external key factors, we propose an algae growth model (AGM) optimized by action dependent heuristic dynamic programming (ADHDP). This model has the structure of information interaction with the outside, which can predict algal bloom with well adaptive ability. In this paper, chlorophyll-a concentration is used as the representative factor of algal bloom. We use ADHDP approach to map the external key factors to the time-varying parameters, so the AGM can be adjusted to realize the self-adaptive prediction with the changes in external environments. Compared with different prediction methods, the simulation result shows that the ADHDP-AGM prediction model can accurately predict the chlorophyll-a concentration under different data distributions. Moreover, the prediction process shows that the time-varying parameters in AGM conform to the evolution trend of chlorophyll-a concentration in fact, which further improves the interpretability of prediction model. It provides a new perspective for building a data-driven prediction model with clear physical significance, and makes the mechanism research and data science further fusion.

The proportional integral and derivative (PID) controller is extensively applied in many applications. However, three parameters must be properly adjusted to ensure effective performance of the control system: the proportional gain (KP), integral gain (KI) and derivative gain (KD). Therefore, the aim of this paper is to optimize and improve the stability, convergence and performance in autotuning the PID parameter by using a deterministic Q-SLP algorithm. The proposed method is a combination of the swarm learning process (SLP) algorithm and Q-learning algorithm. The Q-learning algorithm is applied to optimize the weight updating of the SLP algorithm based on the new deterministic rule and closed-loop stabilization of the learning rate. To validate the global optimization of the deterministic rule, it is proven based on the Bellman equation, and the stability of the learning process is proven with respect to the Lyapunov stability theorem. Additionally, to demonstrate the superiority of the performance and convergence in autotuning the PID parameter, simulation results of the proposed method are compared with those based on the central position control (CPC) system using the traditional SLP algorithm, the whale optimization algorithm (WOA) and improved particle swarm optimization (IPSO). The comparison shows that the proposed method can provide results superior to those of the other algorithms with respect to both performance indices and convergence.

A fusion image combining both anatomical and functional information obtained by registering medical images of two different modalities, Positron Emission Tomography (PET) and Computed Tomography (CT), is of great significance for medical image analysis and diagnosis. Medical image registration relies on similarity measure which is low between PET/CT image voxels and therefore PET/CT registration is a challenging task. To address this issue, this paper presents an unsupervised end-to-end method, DenseRegNet, for deformable 3D PET/CT image registration. The method consists of two stages: (1) predicting 3D displacement vector field (DVF); and (2) registering 3D image. In the 3D DVF prediction stage, a two-level similarity measure together with a deformation regularization is proposed as loss function to optimize network training.In the image registration stage, a resampler and a spatial transformer are utilized to obtain the registration results. In this paper, 663 pairs of Uptake Value (SUV) and Hounsfield Unit (Hu) patches of 106 patients, 227 pairs of SUV and Hu patches of 35 patients and 259 pairs of SUV and Hu patches of 35 patients are randomly selected as training, validation and test set, respectively. Normalized cross correlation (NCC), intersection over union (IoU) of liver bounding box and euclidean distance (ED) on landmark points are used to evaluate the registration results. Experiment results show that the proposed method, DenseRegNet, achieves the best results in terms of liver bounding box IoU and ED, and the second highest value of NCC. For a trained model, given a new pair of PET/CT images, the registration result can be obtained with only one forward calculation within 10 seconds. Through qualitative and quantitative analyses, we demonstrate that, compared with other deep learning registration models, the proposed DenseRegNet achieves improved results in the challenging deformable PET/CT registration task.

Aiming at the problem that the Proportional-Integral-Derivative (PID) control strategy needs to readjust controller parameters for different Parkinson's disease (PD) states. This work proposes an improved control strategy that considers an artificial neural network control scheme. A backpropagation neural network (BPNN) controller is designed to solve the above problem and further to improve the performance of the closed-loop control strategy. The training data set of the BPNN controller is obtained by controlling eight different PD states (PDa - PDh) by the PID controller and the BPNN controller is trained by the training data set to obtain a set of optimal weights. By modulating other different PD states (e.g. PD1 - PD3), the effectiveness of the PID-structure controller and BPNN controller are compared. We find that the BPNN controller can modulate different PD states without changing the controller parameters and reduce energy expenditure by 58.26%. This work is helpful for the design of more effective closed-loop deep brain stimulation (DBS) systems for clinical applications and provides a framework for the further development of closed-loop DBS.

The liver is a common site for the development of primary (i.e., originating from the liver, e.g., hepatocellular carcinoma) or secondary (i.e., spread to the liver, e.g., colorectal cancer) tumor. Due to its complex background, heterogeneous, and diffusive shape, automatic segmentation of tumor remains a challenging task. So far, only the interactive method has been adopted to obtain the acceptable segmentation results of a liver tumor. In this paper, we design an Attention Hybrid Connection Network architecture which combines soft and hard attention mechanism and long and short skip connections. We also propose a cascade network based on the liver localization network, liver segmentation network, and tumor segmentation network to cope with this challenge. Simultaneously, the joint dice loss function is proposed to train the liver localization network to obtain the accurate 3D liver bounding box, and the focal binary cross entropy is used as a loss function to fine-tune the tumor segmentation network for detecting more potentially malignant tumor and reduce false positives. Our framework is trained using the 110 cases in the LiTS dataset and extensively evaluated by the 20 cases in the 3DIRCADb dataset and the 117 cases in the Clinical dataset, which indicates that the proposed method can achieve faster network convergence and accurate semantic segmentation and further demonstrate that the proposed method has a good clinical value.

Scientific literature summarization aims to summarize related papers into a survey. It can help researchers, especially newcomers, to quickly know the current situation of their professional area from massive literature. Since the structure of documents is very important for scientific literature summarization, and it is obviously observed that there is a relationship and semantic information hidden in document structures of scientific literature, therefore, we employ a hierarchical attention model to learn document structure from all the papers for summarization. In particular, we utilize attention mechanism to capture relations among document structure and learn semantic information on document discourse levels, and judge the importance of each sentence according to its surroundings by the information obtained. Moreover, we automatically construct a scientific literature data set consisting of surveys and their references. We evaluate our proposed model on this dataset with ROUGE metrics, experiments prove that our approach is effective, and our model outperforms several baselines.

Person re-identification has become a challenging task due to various factors. One key to effective person re-identification is the extraction of the discriminative features of a person's appearance. Most previous works based on deep learning extract pedestrian characteristics from neural networks but only from the top feature layer. However, the low-layer feature could be more discriminative in certain circumstances. Hence, we propose a method, named the multi-level feature network with multiple losses (MFML), which has a multi-branch network architecture that consists of multiple middle layers and one top layer for feature representations. To extract the discriminative middle-layer features and have a good effect on deeper layers, we utilize the triplet loss function to train the middle-layer features. For the top layer, we focus on learning more discriminative feature representations, so we utilize the hybrid loss (HL) function to train the top-layer feature. Instead of concatenating multilayer features directly, we concatenate the weighted middle-layer features and the weighted top-layer feature as the discriminative features in the testing phase. The extensive evaluations conducted on three datasets show that our method achieves a competitive accuracy level compared with the state-of-the-art methods.

Liver cancer is one of the most common cancers. Liver tumor segmentation is one of the most important steps in treating liver cancer. Accurate tumor segmentation on computed tomography (CT) images is a challenging task due to the variation of the tumor's shape, size, and location. To this end, this paper proposes a liver tumor segmentation method on CT volumes using multi-scale candidate generation method (MCG), 3D fractal residual network (3D FRN), and active contour model (ACM) in a coarse-to-fine manner. First, livers are segmented using 3D U-Net and then MCG is performed on these liver regions for obtaining tumor candidates (all superpixel blocks). Second, 3D FRN is proposed to further determine tumor regions, which is considered as coarse segmentation results. Finally, the ACM is used for tumor segmentation refinement. The proposed 3D MCG-FRN + ACM is trained using the 110 cases in the LiTS dataset and evaluated on a public liver tumor dataset of the 3DIRCADb with dice per case of 0.67. The experimentations and comparisons demonstrate the performance advantage of the 3D MCG-FRN + ACM compared to other segmentation methods.

In the positron emission tomography/computed tomography (PET/CT) image diagnosis report, the semantic analysis of image findings section is an important part of the automatic diagnosis of medical image, which is an essential step for extracting keywords and abnormal sentences in the diagnostic report. To this end, this paper combines visibility attribute extraction network (VAE-Net) and bi-directional gated recurrent unit (BiGRU) into cascade networks to solve the tasks of attribute extraction and anomaly detection. First, a visibility attribute (VA) is defined to summary the vocabulary into 12 patterns based on the language characteristics in image findings. Second, a visibility attribute extraction network (VAE-Net) is developed to automatically extract VA from word embeddings, which is composed of residual convolutional neural network (residual CNN), BiGRU, and conditional random field (CRF). Finally, word embeddings and the corresponding VA are input into BiGRU and softmax to perform sentence-level anomaly detections. We evaluate the proposed method on a proprietary Chinese PET/CT diagnostic report dataset with an F1-score of 94.35% in the attribute extraction, an F1-score of 96.40% in sentence-level anomaly detection, and an F1-score of 96.77% in case-level anomaly detection. Besides, a publicity English national center for biotechnology information (NCBI) disease corpus dataset is used for externed validation with an F1-score of 95.81% in disease detection. The experimental results demonstrate the advantage of the proposed cascade networks as compared to other related methods.

Genetic network programming (GNP) is a relatively new type of graph-based evolutionary algorithm, which designs a directed graph structure for its individual representation. A number of studies have demonstrated its expressive ability to model complicated problems/systems and explored it from the perspectives of methodologies and applications. However, the unique features of its directed graph are relatively unexplored, which cause unnecessary dilemma for the further usage and promotion. This paper is dedicated to uncover this issue systematically and theoretically. It is proved that the traditional GNP with uniform genetic operators does not consider the ``transition by necessity”feature of the directed graph, which brings the unnecessary difficulty of evolution to cause invalid/negative evolution problems. Consequently, simplified genetic operators are developed to address these problems. Experimental results on two benchmark testbeds of the agent control problems are carried out to demonstrate its superiority over the traditional GNP and the state-of-the-art algorithms in terms of fitness results, search speed, and computation time.

In this paper, we introduce a novel and efficient hybrid trajectory planning method for autonomous driving in highly constrained environments. The contributions of this paper are fourfold. First, we present a trajectory planning framework that is able to handle geometry constraints, nonholonomic constraints, and dynamics constraints of cars in a humanlike and layered fashion and generate curvature-continuous, kinodynamically feasible, smooth, and collision-free trajectories in real time. Second, we present a derivative-free global path modification algorithm to extract high-order state information in free space for state sampling. Third, we extend the regular state-space sampling method widely used in on-road autonomous driving systems to a multi-phase deterministic state-space sampling method that is able to approximate complex maneuvers. Fourth, we improve collision checking accuracy and efficiency by using a different car footprint approximation strategy and a two-phase collision checking routine. A range of challenging simulation experiments show that the proposed method returns high-quality trajectories in real time and outperforms existing planners, such as hybrid A* and conjugate-gradient descent path smoother in terms of path quality, efficiency, and computation resources used.