The aim of explainable machine learning is to aid humans in understanding how exactly complex machine learning models work. Machine learning models have offered great performance in various areas. However, the mechanisms behind how the model works and how decisions are being made remain unknown. This specific constraint increases the user’s hesitation to trust the results of the model and even to improve their performance further. Counterfactual explanation is one method to offer explainability in machine learning by indicating what would have happened if the input of a model was modified in a specific way. Fraud is the action of acquiring something from someone else in a dishonest manner. Companies’ and organizations’ vulnerability to malicious actions has been increasing due to the development of digitalization. Machine learning applications have been successfully put in place to tackle fraudulent actions. However, the severity of the impact of fraudulent actions has highlighted the need for further scientific exploration of the topic. The current research will attempt to do so by studying counterfactual explanations related to fraud detection. Latent-CF is a method for counterfactual generation that utilizes an autoencoder and gradient descent in its latent space. LatentCF++ is an extension of Latent-CF. It utilizes a classifier and an autoencoder. The aim is to perturb the encoded latent representation through a gradient descent optimization for counterfactual generation so that the initially undesired class is then classified with the desired prediction. Compared to Latent-CF, LatentCF++ uses Adam optimization and adds further constraints to ensure that the generated counterfactual’s class probability surpasses the set decision boundary. The research question the current thesis addresses is: “To what extent can LatentCF++ provide reliable counterfactual explanations in fraud detection?”. In order to provide an answer to this question, the study is applying an experiment to implement a new application of LatentCF++. The current experiment utilizes a onedimensional convolutional neural network as a classifier and a deep autoencoder for counterfactual generation in fraud data. This study reports satisfying results regarding counterfactual explanation production of LatentCF++ on fraud detection. The classification is quite accurate, while the reconstruction loss of the deep autoencoder employed is very low. The validity of the counterfactual examples produced is lower than the original study while the proximity is lower. Compared to baseline models, k-nearest neighbors outperform LatentCF++ in terms of validity and Feature Gradient Descent in terms of proximity.

MRI is favorable for brain imaging due to its excellent soft tissue contrast and absence of harmful ionizing radiation. Many have proposed supervised multimodal neural networks for automatic brain tumor segmentation and showed promising results. However, they rely on large amounts of labeled data to generalize well. The trained network is also highly specific to the task and input. Missing inputs will most likely have a detrimental effect on the network’s predictions, if it works at all. The aim of this thesis work is to implement a deep neural network that learns the general representation of multimodal MRI images in an unsupervised manner and is insensitive to missing modalities. With the latent representation, labeled data are then used for brain tumor segmentation. A variational autoencoder and an unified representation network are used for repre- sentation learning. Fine-tuning or joint training was used for segmentation task. The performances of the algorithms at the reconstruction task was evaluated using the mean- squared error and at the segmentation task using the Dice coefficient. Both networks demonstrated the possibility in learning brain MR representations, but the unified representation network was more successful at the segmentation task.

The widespread of machine learning and deep learning in commercial and industrial settings has seen a dramatic up-rise. While the traditional software engineering techniques have overlap between machine learning model development, fundamental differences exist which affect both scientific disciplines. The current state-of-the-art argues that most challenges in software engineering of deep learning applications stem from poorly defined software requirements, tightly coupled architecturesand hardware-induced development issues. However the majority of the current work on this topic stems from literature reviews and requires validation in an industrial context. The work aims to validate the findings of the academia through the development of the autoencoder model for gearbox fault detection. The model has been developed as a part of the ongoing campaign from Volvo Construction Equipment towards introducing AI-based solution in quality control and production. Findings of the work are mostly aligned with the current state-of-the-art, where poorly defined software requirements and hardware-induced issues have been experienced, but the tightly-coupled architecture did not characterize the final product. Along with the confirmation of the previous findings, the work presents a recommendation for practitioners of software engineering for deep learning models in the form of technological rule which addresses the hardware-induced issues of development through the contribution of a method for calculating the memory requirements of the model and batch during the training phase. / <p>Collaboration with Volvo Construction Equipment</p>

京都大学 / 新制・論文博士 / 博士(医学) / 乙第13501号 / 論医博第2260号 / 新制||医||1061(附属図書館) / 京都大学大学院医学研究科医学専攻 / (主査)教授 小林 恭, 教授 中本 裕士, 教授 黒田 知宏 / 学位規則第4条第2項該当 / Doctor of Medical Science / Kyoto University / DFAM

Radio map estimation (RME) is the task of predicting radio power at all locations in a two-dimensional area and at all frequencies in a given band. This thesis explores four deep learning approaches to RME: dual path autoencoders, skip connection autoencoders, diffusion, and joint learning with transmitter localization.

Myoelectric control schemes aim to utilize the surface electromyography (EMG) signals which are the electric potentials directly measured from skeletal muscles to control wearable robots such as exoskeletons and prostheses. The main challenge of myoelectric controls is to increase and preserve the signal quality by minimizing the effect of confounding factors such as muscle fatigue or electrode shift. Current research in myoelectric control schemes are developed to work in ideal laboratory conditions, but there is a persistent need to have these control schemes be more robust and work in real-world environments. Following the manifold hypothesis, complexity in the world can be broken down from a high-dimensional space to a lower-dimensional form or representation that can explain how the higher-dimensional real world operates. From this premise, the biological actions and their relevant multimodal signals can be compressed and optimally pertinent when performed in both laboratory and non-laboratory settings once the learned representation or manifold is discovered. This thesis outlines a method that incorporates the use of a contrastive variational autoencoder with an integrated classifier on multimodal sensor data to create a compressed latent space representation that can be used in future myoelectric control schemes.

Mechanical devices in industriy are equipped with numerous sensors to capture thehealth state of the machines. The reliability of the machine’s health system depends on thequality of sensor data. In order to predict the health state of sensors, abnormal behaviourof sensors must be detected to avoid unnecessary cost.We proposed LSTM autoencoder in which the objective is to reconstruct input time seriesand predict the next time instance based on historical data, and we evaluate anomaliesin multivariate time series via reconstructed error. We also used exponential moving averageas a preprocessing step to smooth the trend of time series to remove high frequencynoise and low frequency deviation in multivariate time series data.Our experiment results, based on different datasets of multivariate time series of gasturbines, demonstrate that the proposed model works well for injected anomalies and realworld data to detect the anomaly. The accuracy of the model under 5 percent infectedanomalies is 98.45%.

Dimensionality reduction techniques (DR) enhance data interpretability and reduce space complexity, though at the cost of information loss. Such methods have been prevalent in the Structural Health Monitoring (SHM) anomaly detection literature. While DR is favorable in supervised anomaly detection, where possible novelties are known a priori, the efficacy is less clear in unsupervised detection. In this work, we perform a detailed assessment of the DR performance trade-offs to determine whether the information loss imposed by DR can impact SHM performance for previously unseen novelties. As a basis for our analysis, we rely on an SHM anomaly detection method operating on input signals' fast Fourier transform (FFT). FFT is regarded as a raw, frequency-domain feature that allows studying various DR techniques. We design extensive experiments comparing various DR techniques, including neural autoencoder models, to capture the impact on two SHM benchmark datasets exclusively. Results imply the loss of information to be more detrimental, reducing the novelty detection accuracy by up to 60\% with autoencoder-based DR. Regularization can alleviate some of the challenges though unpredictable. Dimensions of substantial vibrational information mostly survive DR; thus, the regularization impact suggests that these dimensions are not reliable damage-sensitive features regarding unseen faults. Consequently, we argue that designing new SHM anomaly detection methods that can work with high-dimensional raw features is a necessary research direction and present open challenges and future directions. / M.S. / Structural health monitoring (SHM) aids the timely maintenance of infrastructures, saving human lives and natural resources. Infrastructure will undergo unseen damages in the future. Thus, data-driven SHM techniques for handling unlabeled data (i.e., unsupervised learning) are suitable for real-world usage. Lacking labels and defined data classes, data instances are categorized through similarities, i.e., distances. Still, distance metrics in high-dimensional spaces can become meaningless. As a result, applying methods to reduce data dimensions is currently practiced, yet, at the cost of information loss. Naturally, a trade-off exists between the loss of information and the increased interpretability of low-dimensional spaces induced by dimensionality reduction procedures. This study proposes an unsupervised SHM technique that works with low and high-dimensional data to assess that trade-off. Results show the negative impacts of dimensionality reduction to be more severe than its benefits. Developing unsupervised SHM methods with raw data is thus encouraged for real-world applications.

Unsupervised anomaly detection is the task of identifying examples that differ from the normal or expected pattern without the use of labeled training data. Our research addresses shortcomings in two existing anomaly detection algorithms, Kernel Principal Component Analysis (KPCA) and Autoencoders (AE), and proposes novel solutions to improve both of their performances in the unsupervised settings. Anomaly detection has several useful applications, such as intrusion detection, fault monitoring, and vision processing. More specifically, anomaly detection can be used in autonomous driving to identify obscured signage or to monitor intersections. Kernel techniques are desirable because of their ability to model highly non-linear patterns, but they are limited in the unsupervised setting due to their sensitivity of parameter choices and the absence of a validation step. Additionally, conventionally KPCA suffers from a quadratic time and memory complexity in the construction of the gram matrix and a cubic time complexity in its eigendecomposition. The problem of tuning the Gaussian kernel parameter, $sigma$, is solved using the mini-batch stochastic gradient descent (SGD) optimization of a loss function that maximizes the dispersion of the kernel matrix entries. Secondly, the computational time is greatly reduced, while still maintaining high accuracy by using an ensemble of small, textit{skeleton} models and combining their scores. The performance of traditional machine learning approaches to anomaly detection plateaus as the volume and complexity of data increases. Deep anomaly detection (DAD) involves the applications of multilayer artificial neural networks to identify anomalous examples. AEs are fundamental to most DAD approaches. Conventional AEs rely on the assumption that a trained network will learn to reconstruct normal examples better than anomalous ones. In practice however, given sufficient capacity and training time, an AE will generalize to reconstruct even very rare examples. Three methods are introduced to more reliably train AEs for unsupervised anomaly detection: Cumulative Error Scoring (CES) leverages the entire history of training errors to minimize the importance of early stopping and Percentile Loss (PL) training aims to prevent anomalous examples from contributing to parameter updates. Lastly, early stopping via Knee detection aims to limit the risk of over training. Ultimately, the two new modified proposed methods of this research, Unsupervised Ensemble KPCA (UE-KPCA) and the modified training and scoring AE (MTS-AE), demonstrates improved detection performance and reliability compared to many baseline algorithms across a number of benchmark datasets. / Master of Science / Anomaly detection is the task of identifying examples that differ from the normal or expected pattern. The challenge of unsupervised anomaly detection is distinguishing normal and anomalous data without the use of labeled examples to demonstrate their differences. This thesis addresses shortcomings in two anomaly detection algorithms, Kernel Principal Component Analysis (KPCA) and Autoencoders (AE) and proposes new solutions to apply them in the unsupervised setting. Ultimately, the two modified methods, Unsupervised Ensemble KPCA (UE-KPCA) and the Modified Training and Scoring AE (MTS-AE), demonstrates improved detection performance and reliability compared to many baseline algorithms across a number of benchmark datasets.

This work aims to develop a deep learning model for land cover quantification through hyperspectral unmixing using an unsupervised autoencoder. Land cover identification and classification is instrumental in urban planning, environmental monitoring and land management. With the technological advancements in remote sensing, hyperspectral imagery which captures high resolution images of the earth's surface across hundreds of wavelength bands, is becoming increasingly popular. The high spectral information in these images can be analyzed to identify the various target materials present in the image scene based on their unique reflectance patterns. An autoencoder is a deep learning model that can perform spectral unmixing by decomposing the complex image spectra into its constituent materials and estimating their abundance compositions. The advantage of using this technique for land cover quantification is that it is completely unsupervised and eliminates the need for labelled data which generally requires years of field survey and formulation of detailed maps. We evaluate the performance of the autoencoder on various synthetic and real hyperspectral images consisting of different land covers using similarity metrics and abundance maps. The scalability of the technique with respect to landscapes is assessed by evaluating its performance on hyperspectral images spanning across 100m x 100m, 200m x 200m, 1000m x 1000m, 4000m x 4000m and 5000m x 5000m regions. Finally, we analyze the performance of this technique by comparing it to several supervised learning methods like Support Vector Machine (SVM), Random Forest (RF) and multilayer perceptron using F1-score, Precision and Recall metrics and other unsupervised techniques like K-Means, N-Findr, and VCA using cosine similarity, mean square error and estimated abundances. The land cover classification obtained using this technique is compared to the existing United States National Land Cover Database (NLCD) classification standard. / Master of Science / This work aims to develop an automated deep learning model for identifying and estimating the composition of the different land covers in a region using hyperspectral remote sensing imagery. With the technological advancements in remote sensing, hyperspectral imagery which captures high resolution images of the earth's surface across hundreds of wavelength bands, is becoming increasingly popular. As every surface has a unique reflectance pattern, the high spectral information contained in these images can be analyzed to identify the various target materials present in the image scene. An autoencoder is a deep learning model that can perform spectral unmixing by decomposing the complex image spectra into its constituent materials and estimate their percent compositions. The advantage of this method in land cover quantification is that it is an unsupervised technique which does not require labelled data which generally requires years of field survey and formulation of detailed maps. The performance of this technique is evaluated on various synthetic and real hyperspectral datasets consisting of different land covers. We assess the scalability of the model by evaluating its performance on images of different sizes spanning over a few hundred square meters to thousands of square meters. Finally, we compare the performance of the autoencoder based approach with other supervised and unsupervised deep learning techniques and with the current land cover classification standard.