Electroencephalogram (EEG) signal has been widely applied in detecting human emotion. Individual differences limit the generalization in cross-subject classification since the release of emotion will be definitely different across persons even with the same emotion stimuli. Previous research utilizes domain adaptation to solve this problem in a leave-one-subject-out training that is learning common emotion-related features from many subjects and testing on a new subject, which requires abundant labeled data. This paper proposed a novel one-to-one domain adaptation method, the Task-specific Domain Adversarial Neural Network (T-DANN) which transfers knowledge from either one subject to predict on another subject or knowledge from one phase to predict on another phase within the same subject. Therefore, T-DANN is more flexible and requires much less data during training. T-DANN is an adversarial training method which adapts the conditional distribution between domains and adapts classification boundaries between classes simultaneously. Compared with data from different subjects, phase data from the same subject has much deeper correlation thus enhances the prediction of emotion in new phase. We evaluated our method on EEG emotion benchmark dataset SEED. The experiments showed that our proposed method outperformed other baseline methods in cross-subject adaptation. By cross-phase adaptation, our method achieved accuracy that approximately 3% lower than the state-of-the-art method but only used 1/14 labeled data, indicating the priority of our proposed method in real-time application.

Recently, researchers have utilized neural networks to accurately solve partial differential equations (PDEs), enabling the mesh-free method for scientific computation. Unfortunately, the network performance drops when encountering a high nonlinearity domain. To improve the generalizability, we introduce the novel approach of employing multi-task learning techniques, the uncertainty-weighting loss and the gradients surgery, in the context of learning PDE solutions. The multi-task scheme exploits the benefits of learning shared representations, controlled by cross-stitch modules, between multiple related PDEs, which are obtainable by varying the PDE parameterization coefficients, to generalize better on the original PDE. Encouraging the network pay closer attention to the high nonlinearity domain regions that are more challenging to learn, we also propose adversarial training for generating supplementary high-loss samples, similarly distributed to the original training distribution. In the experiments, our proposed methods are found to be effective and reduce the error on the unseen data points as compared to the previous approaches in various PDE examples, including high-dimensional stochastic PDEs.

In order to improve Remaining Useful Life (RUL) prediction accuracy for rolling bearings under defect progressing, robustness for individual difference and fluctuation of vibration features are challenging issues. In this research, we propose a novel RUL prediction method that uses a hierarchical Bayesian method to consider the individual difference of RUL, and uses an intermediate variable indicating the defect condition instead of predicting RUL directly from vibration features. The proposed method can perform a monotonous RUL prediction curve and improved prediction accuracy especially for early stage of defect progression.

Metric learning algorithms create distance metrics to capture the important relationships among features. These algorithms have been successful in low dimensional data. However, it is a challenge to handle high dimensional problems in real-world applications. This paper applies in high dimensional clustering problems a Reinforcement Learning (RL) based metric filtering approach in Evolutionary Distance Metric Learning (EDML), which relies on an evolutionary approach to optimize its distance metric. The proposed framework called High Dimensional Reinforced (HDR)-EDML has two novelties. The first is using a function approximation RL method called Least-Squares Policy Iteration (LSPI) to create a feature selection control strategy in high-dimensional input space to filter the metric. The second is adopting a two-way information exchange approach between LSPI and EDML Evolutionary Algorithm (EA) in metric learning. In the first way, LSPI will learn and send feedback that will affect the EDML metric creation. In the second way, the evolutionary feature prioritizing of EDML is utilized by LSPI in its learning process. These two novelties aim to reduce the number of selected features while maintaining the clustering performance. HDR-EDML is compared to conventional K-means, Information-Theoretic Metric Learning (ITML), normal EDML, and EDML with l1norm regularization for feature selection as baselines. Moreover, 3 different HDR-EDML approaches are examined to explore different ways of information exchange. Results show a significant decrease in the number of features while maintaining accuracy as well as reduced computational time and memory compared to another RL-based filtering method.

We proposed a method to extract causal relations of clusters from multi-dimensional event sequence data, along with another method to detect the changes of the extracted relations over time. The proposed Granger Cluster Sequence Mining (GCSM) algorithm identifies the pairs of spatial data clusters that have causality over time with each other. It extends the Cluster Sequence Mining algorithm, which utilized a statistical inference technique to identify occurrence relation, with a causality based on Granger causality. We also proposed a statistical model to infer the changes over time of each extracted causal relation. With experiments using synthetic data and semi-real data, we confirmed that the algorithm works correctly, and able to extract the embedded causal relations with high F-score and high accuracy of change points.

In this paper, we propose two modified neural networks based on dual path multi-scale fusion networks (SFANet) and SegNet for accurate and efficient crowd counting. Inspired by SFANet, the first model, which is named M-SFANet, is attached with atrous spatial pyramid pooling (ASPP) and context-aware module (CAN). The encoder of M-SFANet is enhanced with ASPP containing parallel atrous convolutional layers with different sampling rates and hence able to extract multi-scale features of the target object and incorporate larger context. To further deal with scale variation throughout an input image, we leverage the CAN module which adaptively encodes the scales of the contextual information. The combination yields an effective model for counting in both dense and sparse crowd scenes. Based on the SFANet decoder structure, M-SFANet`s decoder has dual paths, for density map and attention map generation. The second model is called M-SegNet, which is produced by replacing the bilinear upsampling in SFANet with max unpooling that is used in SegNet. This change provides a faster model while providing competitive counting performance. Designed for high-speed surveillance applications, M-SegNet has no additional multi-scale-aware module in order to not increase the complexity. Both models are encoder-decoder based architectures and are end-to-end trainable. We conduct extensive experiments on five crowd counting datasets and one vehicle counting dataset to show that these modifications yield algorithms that could improve state-of-the-art crowd counting methods. Codes are available at https://github.com/Pongpisit-Thanasutives/Variations-of-SFANet-for-Crowd-Counting.

This paper proposed a method to detect changes in causal relations over a multi-dimensional sequence of events. Cluster Sequence Mining algorithm was modified to extract causal relations in the form of g-patterns: a pair of clusters of events that have their occurrence time determined by Granger causality. This paper also proposed the pattern time signature, a probabilistic density function of the cluster sequence occurring at any given time. Synthetic data were used for validation. The result shows that the proposed algorithm can correctly identify the changes in causal relations even under noisy data.

Although numerous systems for sleep quality assessment are available on the market nowadays, they usually fail due to two main reasons. First, their sleep quality assessment is not based on scientific evidence. Second, many of these systems require additional hardware to identify the sleep state, thus becoming expensive and impractical. We address the first problem by proposing a novel concept, called SleepAge, which considers the relationship between age and sleep quality and then estimates the sleep pattern to evaluate the sleep quality. This concept relies on age because it is generally available from health records and is an objective parameter that can be easily updated. As sounds such as snoring, bruxism, and body movements can reflect sleep patterns, we address the second problem by recording these sounds using a smartphone. We also apply a novel method for data collection that considers participants from a wide range of ages over long recording periods in home environments. Therefore, we can model normal sleep patterns for different age groups. Moreover, we estimate SleepAge from the large repository of sound dataset using various deep learning models. To this end, sound data are represented as spectrograms, and state-of-the-art deep learning models are used for SleepAge estimation. Experimental results indicate accurate estimation, and a network pretrained on different images further enhances the estimation accuracy of unseen data.

The performance of human workers can be fluctuated due to changes in the cognitive state during sustained work. Though past researches have made human performance monitoring possible by utilizing physiological signals, little attention has been paid to the context of office works. This research proposes a multimodal approach to estimate office task performance. A transcription typing experiment was conducted to simulate the real working environment while typing speed and error rate represented as performance metrics. Physiological data collected during the experiment, together with conventional machine learning algorithms showed feasibility to accurately predict two levels (good/bad) of task performance. More importantly, a comprehensive comparison between choices of modality suggests that using data from particular sources could gain predictive performance comparable to the complete set of modalities.

Data collection plays an important role in business agility; data can prove valuable and provide insights for important features. However, conventional data collection methods can be costly and time-consuming. This paper proposes a hybrid system R-EDML that combines a sequential feature selection performed by Reinforcement Learning (RL) with the evolutionary feature prioritization of Evolutionary Distance Metric Learning (EDML) in a clustering process. The goal is to reduce the features while maintaining or increasing the accuracy leading to less time complexity and future data collection time and cost reduction. In this method, features represented by the diagonal elements of EDML matrices are prioritized using a differential evolution algorithm. Further, a selection control strategy using RL is learned by sequentially inserting and evaluating the prioritized elements. The outcome offers the best accuracy R-EDML matrix with the least number of elements. Diagonal R-EDML focusing on the diagonal elements is compared with EDML and conventional feature selection. Full Matrix R-EDML focusing on the diagonal and non-diagonal elements is tested and compared with Information-Theoretic Metric Learning. Moreover, R-EDML policy is tested for each EDML generation and across all generations. Results show a significant decrease in the number of features while maintaining or increasing accuracy.

In this research, we combine relational learning with multi-domain to develop a formal framework for a recommendation system. The design of our framework aims at: (i) constructing general rules for recommendations, (ii) providing suggested items with clear and understandable explanations, (iii) delivering a broad range of recommendations including novel and unexpected items. We use relational learning to find all possible relations, including novel relations, and to form the general rules for recommendations. Each rule is represented in relational logic, a formal language, associating with probability. The rules are used to suggest the items, in any domain, to the user whose preferences or other properties satisfy the conditions of the rule. The information described by the rule serves as an explanation for the suggested item. It states clearly why the items are chosen for the users. The explanation is in if-then logical format which is unambiguous, less redundant and more concise compared to a natural language used in other explanation recommendation systems. The explanation itself can help persuade the user to try out the suggested items, and the associated probability can drive the user to make a decision easier and faster with more confidence. Incorporating information or knowledge from multiple domains allows us to broaden our search space and provides us with more opportunities to discover items which are previously unseen or surprised to a user resulting in a wide range of recommendations. The experiment results show that our proposed algorithm is very promising. Although the quality of recommendations provided by our framework is moderate, our framework does produce interesting recommendations not found in the primitive single-domain based system and with simple and understandable explanations.

In this paper, we focus on a method that integrates a physical model into a neural network. This study proposes a neural network that can predict two components, namely outputs based on a physical model and its model discrepancy. To achieve such a goal, we propose a novel neural network architecture and associated loss functions designed based on a target physical model. The physical model is used as a regularizer of spatial behavior where output from the neural network is used as an intermediate variable. Then, the model discrepancy is defined as its residual to the observation value. We also propose a network architecture which has Shared and Non-Shared networks, and the neural network can be trained by alternate optimization. We constructed the proposed method with wind prediction in the upper troposphere based on thermal wind equations as an example. The experimental results demonstrate that the proposed method can achieve higher predictive accuracy than normal convolutional neural network or using thermal wind equation, also the obtained model discrepancy expresses convergence and divergence of wind vectors.

This study proposes a novel distance metric learning method called evolutionary distance metric learning (EDML) to improve clustering quality that simultaneously evaluates inter- and intra-clusters. While we also provide an extension which integrates kernelization technique to the proposed method n amely kernelized evolutionary distance metric learning (K-EDML). Hence, the non-linear transformation of distance metric can be performed while maintaining all properties of EDML. The proposed methods are able to handle either class label or pairwise constraints and directly improve any clustering index as an objective function. Both can be viewed as utilization of cluster-level soft constraints, unlike other instance-level hard constraints which sometimes collapse the clustering. Also, maintaining neighbor relation of clusters can lead to better visualization of the clustering result. For multimodality problem of the objective function, an evolutionary algorithm (EA), differential evolution with self-adapting control parameters and generalized opposition-based learning (GOjDE), is employed to optimize a metric transform matrix based on the Mahalanobis distance. We empirically demonstrate the drawback of EDML in non-linearly separable input space and illustrate the benefit of kernel function to extension K-EDML method by showing its superior result benefits to other clustering algorithms in the semi-supervised clustering on various real-world datasets.

We propose a novel mining algorithm called cluster sequence mining (CSM) to extract cluster pairs with occurrence correlation from event sequence data. CSM extracts patterns with a pair of clusters that satisfies space proximity of the individual clusters and temporal proximity between events from different clusters in time intervals. CSM extends a unique co-occurring cluster mining (CCM) algorithm by considering the order of event occurrences and distribution of time intervals. The probability density of time intervals is inferred using Bayesian inference for robustness against uncertainty. To improve inference accuracy of the density function of time intervals, we utilize the idea of dynamic programming (DP) matching to obtain the correspondence between multiple event occurrences. With an experiment using synthetic data, we confirm that CSM is capable of extracting clusters with a high F-measure and low estimation error of the time interval distribution even under uncertainty. In addition, we find that DP matching can improve the inference accuracy of the density function of time intervals. Finally, CSM is applied to a real-world acoustic emission event sequence data set to evaluate damage interactions in a fuel cell.

We propose a method to generate explainable recommendation rules on cross-domain problems. Our two main contributions are: i) using relational learning to generate the rules which are able to explain clearly why the items were recommended to the particular user, ii) using the user's preferences of items on different domains and item attributes to generate novel or unexpected recommendations for the user. To illustrate that our method is indeed feasible and applicable, we conducted experiments on music and movie domains.

Many research studies on distance metric learning (DML) reiterate that the definition of distance between two data points substantially affects clustering tasks. Recently, variety of DML methods have been proposed to improve the accuracy of clustering by learning a distance metric; however, most of them only perform a linear transformation, which yields insignificant to non-linear separable data. This study proposes a DML method which provides an integration of kernelization technique with Mahalanobis-based DML. Thus, non-linear transformation of the distance metric can be performed. Moreover, a cluster validity index is optimized by an evolutionary algorithm. The empirical results on semi-supervised clustering suggest the promising result on both synthetic and real-world data set.

Multimodality has been recently exploited to overcome the challenges of emotion recognition. In this paper, we present a study of fusion of electroencephalogram (EEG) features and musical features extracted from musical stimuli at decision level in recognizing the time-varying binary classes of arousal and valence. Our empirical results demonstrate that EEG modality was suffered from the non-stability of EEG signals, yet fusing with music modality could alleviate the issue and enhance the performance of emotion recognition.

ART-2a has been shown to be effective against stream data clustering with unknown number of cluster in nature. As data grows, ART-2a running time become a major problem. We proposed a new algorithm, ART-2b, whose runtime performance is linear to the number of input instances, while still maintaining similar clustering result to ART-2a.

Data stream classifiers that can withstand unusual phenomena in an evolving data stream, such as concept drift and concept evolution, are highly desirable for data stream mining. Most existing methods deal with such phenomena in a supervised manner, which is costly in a real-world scenario. To address this shortcoming, we propose a concept drift detection approach that combines our approach with a semi-supervised adaptive incremental neural gas (A2ING) classifier. Our approach makes use of A2ING's graph topology structure to detect changes in a data stream. We derive a graph's instability around its decision boundary and find the difference in prior and posterior distributions of the criteria. The empirical results show the effectiveness of our method. The classifier requires a relatively low number of true labels compared to existing approaches and shows high effectiveness in change detection.

In our society, people engage in a variety of interactions. To analyze such interactions, we consider these interactions as a game and people as agents equipped with reinforcement learning algorithms. Reinforcement learning algorithms are widely studied with a goal of identifying strategies of gaining large payoffs in games; however, existing algorithms learn slowly because they require a large number of interactions. In this work, we constructed an algorithm that both learns quickly and maximizes payoffs in various repeated games. Our proposed algorithm combines two different algorithms that are used in the early and later stages of our algorithm. We conducted experiments in which our proposed agents played ten kinds of games in self-play and with other agents. Results showed that our proposed algorithm learned more quickly than existing algorithms and gained sufficiently large payoffs in nine games.

In the ocean around the world, depth series ocean data of temperature and salinity have being measured. However, it is difficult to discriminate the errors from the normal data since the variation of ocean areas are different. In this research, using hierarchical clustering, we partitioned the ocean into some areas so that the ocean data have the same variation in each area. Then, transforming the ocean data into sets of sliding windows in consideration of depth series, we applied some anomaly detection methodologies. Finally, we succeeded in assigning high anomaly scores on errors that seemed to be normal.

Estimating emotional states in music listening based on electroencephalogram (EEG) has been capturing the attention of researchers in the past decade. Although deep belief network (DBN) has witnessed the success in various domains including early works in emotion recognition based on EEG, it remains unclear whether DBN could improve emotion classification in music domains, especially in dynamic strategy that considers time-varying characteristics of emotion. This paper presents an early study of applying DBNs to improve emotion recognition in music listening where emotions were annotated continuously in time by subjects. Our subject-dependent results using stratified 10-fold cross-validation strategy suggested that DBNs could improve performance in valence classification with fractal dimension (FD), power spectral density (PSD), and discrete wavelet transform (DWT) features and improve performance in arousal classification with FD and DWT features. Furthermore, we found that the size of sliding window affected classification accuracies when using features in time (FD) and time-frequency (DWT) domains, while smaller window (1-4 seconds) could achieve higher performance compared with a larger window (5-8 seconds).

As emotional responses of a human to stimuli could evolve over the course of time, continuous emotion reporting is essential for the construction of a computational model to capture the temporal evolution of the human emotions. However, continuous emotion assessment is confronting various challenges, especially when using the continuous arousal-valence space. Manipulating emotion annotation data prior to performing emotion recognition is, therefore, necessary. In this paper, we present a study of applying three different signal filtering techniques to smooth annotation data; moving average filter, Savitzky-Golay filter, and median filter. We performed experiments of arousal and valence recognition in music listening tasks employing signals from electroencephalogram (EEG). Fractal dimension approach was adopted to extract informative features from brain dynamics and emotional states were then derived by classification and regression techniques. Our empirical results suggested the promise of the moving average filter that could enhance the performance of emotion classifying and tracking.

To gain a better understanding of how neural ensembles communicate and process information, neural decoding algorithms are used to extract information encoded in their spiking activity. Bayesian decoding is one of the most used neural population decoding approaches to extract information from the ensemble spiking activity of rat hippocampal neurons. Recently it has been shown how Bayesian decoding can be implemented without the intermediate step of sorting spike waveforms into groups of single units. Here we extend the approach in order to make it suitable for online encoding/decoding scenarios that require real-time decoding such as brain-machine interfaces. We propose an online algorithm for the Bayesian decoding that reduces the time required for decoding neural populations, resulting in a real-time capable decoding framework. More specifically, we improve the speed of the probability density estimation step, which is the most essential and the most expensive computation of the spike-sorting-less decoding process, by developing a kernel density compression algorithm. In contrary to existing online kernel compression techniques, rather than optimizing for the minimum estimation error caused by kernels compression, the proposed method compresses kernels on the basis of the distance between the merging component and its most similar neighbor. Thus, without costly optimization, the proposed method has very low compression latency with a small and manageable estimation error. In addition, the proposed bandwidth matching method for Gaussian kernels merging has an interesting mathematical property whereby optimization in the estimation of the probability density function can be performed efficiently, resulting in a faster decoding speed. We successfully applied the proposed kernel compression algorithm to the Bayesian decoding framework to reconstruct positions of a freely moving rat from hippocampal unsorted spikes, with significant improvements in the decoding speed and acceptable decoding error.

In data mining and machine learning, the definition of the distance between two data points substantially affects clustering and classification tasks. We propose a distance metric learning (DML) method for multi-label clustering, that uses evolutionary multi-objective optimization and a cluster validity measure with a neighbor relation that simultaneously evaluates inter- and intra-clusters. The proposed method produces clustering results considering multiple class labels and allows the induction of knowledge regarding relations between class labels in multi-label clustering or between objective functions and elements in transform matrix. Experimental results have shown that the proposed DML method produces better transform matrices than single-objective optimization and is helpful in finding the attributes that affect the trade-off relationship among objective functions.

Globally-covered ocean monitoring system Argo with more than 3,600 small and light-weight drifting buoys is always working for oceanic temperature and salinity measurement. The accumulated big ocean observation data helps many studies such as investigation into climate change mechanism. Although human experts visually confirm and revise quality control (QC) labels, it is difficult to regularize the quality of the ocean observation data of all over the world. Therefore, this paper proposes a method for error detection in Argo observation data to realize an automatic QC with high accuracy equal to human experts. The target dataset is imbalanced data and requires consideration of sequence of both features and quality labels for accurate labeling in each depth. The proposed method utilizes Conditional Random Field (CRF) to assign quality labels for observed temperature and salinity values, and adopts Support Vector Machine (SVM) to design a feature function for numerical attributes. Experimental results have shown that the proposed method showed better accuracy of QC label assignments than those of point-wise prediction method using SVM and the actually operated system in Argo project.

Self-reflection and self-evaluation are effective processes for identifying good learning behavior. These are essential in self-directed learning scenarios because students have to be responsible for their own learning. Although students benefit from doing fine-grained analysis of their own behavior, which we observed in our previous work, asking them to perform tasks such as analysis and making annotations are tedious and take significant amount of time and effort. In this paper, we present our work on the development of incremental affect models that can be used to minimize effort in analyzing and annotating behavior. Incremental models have an added benefit of adaptability to new information, which can be used by future systems to provide up-to-date affect-related feedback in real time.

This study proposes a distance metric learning method based on a clustering index with neighbor relation that simultaneously evaluates inter-and intra-clusters. Our proposed method optimizes a distance transform matrix based on the Mahalanobis distance by utilizing a self-adaptive differential evolution (jDE) algorithm. Our approach directly improves various clustering indices and in principle requires less auxiliary information compared to conventional metric learning methods. We experimentally validated the search efficiency of jDE and the generalization performance.

The present work developed a basis to explore numerous damage events utilizing Self-Organizing Map (SOM) introducing Kullback-Leibler (KL) divergence as an appropriate similarity for frequency spectra of damage events. Firstly, we validated the use of KL divergence to frequency spectra of damage events. The experiment using the datasets of damage related sounds showed that the kernel SOM using KL kernel generates accurate cluster map compared to using general kernel functions and the standard SOM. Afterward, we demonstrated our approach can clarify damage process of Solid Oxide Fuel Cells (SOFC) from acoustic emission (AE) events observed by damage test of SOFC. The damage process was inferred by occurrence frequency of AE events upon the cluster map of SOM, where the occurrence density change was obtained by kernel density estimation (KDE). The presented approach can be a common foundation for the domain experts to clarify fracture mechanism of SOFC and/or to monitor SOFC operation.

Data-centric affect modeling may render itself restrictive in practical applications for three reasons, namely, it falls short of feature optimization, infers discrete affect classes, and deals with relatively small to average sized datasets. Though it seems practical to use the feature combinations already associated to commonly investigated sensors, there may be other potentially optimal features that can lead to new relations. Secondly, although it seems more realistic to view affect as continuous, it requires using continuous labels that will increase the difficulty of modeling. Lastly, although a large scale dataset reflects a more precise range of values for any given feature, it severely hinders computational efficiency. We address these problems when inferring physiology-affect relations from datasets that contain 2-3 million feature vectors, each with 49 features and labelled with continuous affect values. We employ automatic feature selection to acquire near optimal feature subsets and a fast approximate kNN algorithm to solve the regression problem and cope with the challenge of a large scale dataset. Our results show that high estimation accuracy may be achieved even when the selected feature subset is only about 7% of the original features. May the results here motivate the HCI community to pursue affect modeling without being deterred by large datasets and further the discussions on acquiring optimal features for accurate continuous affect approximation.

A crucial issue inputting solid oxide fuel cells (SOFCs) into practical use is the establishment of a technique for evaluating the deterioration of SOFCs. We attempted to capture fracture dynamics measured by acoustic emission (AE) method, employing the following approaches: (1) detection of AE waves utilizing burst extraction method, (2) clustering of AE waves based on the burst level to identify AE types, and (3) visualization of fracture dynamics using adopted self-organizing map. We empirically validated our approach, and obtained a map that can interpret fracture dynamics as physical phenomenon.