Most of the Generation IV nuclear reactors designs are intended to operate with a fast neutron spectrum. This necessitates further investigation into nuclear fuel behaviour because fast neutrons yield a higher neutron multiplicity with fission fragments, significantly impacting the criticality assessment of these reactors. Current nuclear research conducted at the VElocity foR Direct particle Identification spectrometer (VERDI) is focused on enhancing our understanding of the relationship between fission fragment excitation energy and the compound nucleus, as well as the correlation between excitation energy and neutron multiplicity. This is achieved through the innovative double energy, double-velocity technique (2E-2v). However, time-of-flight measurements face challenges due to the inherent Plasma Delay Time (PDT) resulting from the interaction between ions and Passivated Implanted Planar Silicon (PIPS) detectors. The interacting between fission fragments and the PIPS detectors results plasma, which in turn creates a disturbance in the electric field. This disturbance causes there to be a delay in the electron-holes moving across the surface, resulting in a delay in the timing signal, the PDT.  The purpose is of this study is to parameterize the PDT and explore the amount of information that can be extracted from the pulse shapes of heavy ions using Deep Neural Networks. Moreover, to delve deeper into this phenomenon, an experiment was conducted at Institut Laue-Langevin (LOHENGRIN), and the data from this experiment serves as the foundation for this study. This research demonstrates that it is possible to achieve an accuracy surpassing the timing resolution available at VERDI using data from a single detector. The developed neural network exhibits robustness, translational invariance, and the ability to generalize well across all ion signals. Interestingly, when trained solely on the pulse shape, excluding pulse height information, the network still attains highly satisfactory accuracy. Indicating that the pulse shape already holds ample information. However, further investigations are necessary to enhance the network's ability to generalize across data from multiple detectors.

The availability of powerful GPUs and the consequent development of deep neural networks, have brought remarkable results in videogame levels generation, image-to-image translation , video-to-video translation, image inpainting and video generation. Nonetheless, in conditional or constrained settings, unconditioned generative models still suffer because they have little to none control over the generated output. This leads to problems in some scenarios, such as structured objects generation or multimedia manipulation. In the manner, unconstrained GANs fail to generate objects that must satisfy hard constraints (e.g., molecules must be chemically valid or game levels must be playable). In the latter, the manipulation of complex scenes is a challenging and unsolved task, since these scenes are composed of objects and background of different classes. In this thesis , we focus on these two scenarios and propose different techniques to improve deep generative models. First, we introduce Constrained Adversarial Networks (CANs), an extension of GANs in which the constraints are embedded into the model during training. Then we focus on developing novel deep learning models to alter complex urban scenes. In particular, we aim to alter the scene by: i) studying how to better leverage the semantic and instance segmentation to model its content and structure; ii) modifying, inserting and/or removing specific object instances coherently to its semantic; iii) generating coherent and realistic videos where users can alter the object’s position.

This dissertation investigates the problem of representation discovery in discrete Markov decision processes, namely how agents can simultaneously learn representation and optimal control. Previous work on function approximation techniques for MDPs largely employed hand-engineered basis functions. In this dissertation, we explore approaches to automatically construct these basis functions and demonstrate that automatically constructed basis functions significantly outperform more traditional, hand-engineered approaches. We specifically examine two problems: how to automatically build representations for action-value functions by explicitly incorporating actions into a representation, and how representations can be automatically constructed by exploiting a pre-specified task hierarchy. We first introduce a technique for learning basis functions directly in state-action space. The approach constructs basis functions using spectral analysis of a state-action graph which captures the underlying structure of the state-action space of the MDP. We describe two approaches to constructing these graphs and evaluate the approach on MDPs with discrete state and action spaces. We show how our approach can be used to approximate state-action value functions when the agent has access to macro-actions: actions that take more than one time step and have predefined policies. We describe how the state-action graphs can be modified to incorporate information about the macro-actions and experimentally evaluate this approach for SMDPs with discrete state and action spaces. Finally, we describe how hierarchical reinforcement learning can be used to scale up automatic basis function construction. We extend automatic basis function construction techniques to multi-level task hierarchies and describe how basis function construction can exploit the value function decomposition given by a fixed task hierarchy. We demonstrate that combining task hierarchies with automatic basis function construction allows basis function techniques to scale to larger problems and leads to a significant speed-up in learning.

Text documents are generally accompanied by non-textual information, such as authors, dates, publication sources, and, increasingly, automatically recognized named entities. Work in text analysis has often involved predicting these non-text values based on text data for tasks such as document classification and author identification. This thesis considers the opposite problem: predicting the textual content of documents based on non-text data. In this work I study several regression-based methods for estimating the influence of specific metadata elements in determining the content of text documents. Such topic regression methods allow users of document collections to test hypotheses about the underlying environments that produced those documents.

Myocardial infarction is the primary diagnosis to rule out in emergency department chest pain patients. In this retrospective, multi-site study, we compared logistic regression (LR) with machine learning (ML) in predicting which patients were at risk of major adverse cardiac events (MACE) and abnormal troponin. Of the 1,538 patients identified over 43 days, 1,014 were retained of whom 70 suffered a MACE. LR and ML models for MACE were internally validated and achieved similar area under curve (AUC): 0.89 (95% CI: 0.87, 0.93) and 0.92 (95% CI: 0.89, 0.94) respectively. Abnormal troponin models had overlapping AUCs. Two novel clinical decision scores were derived: the Preliminary Chest Pain Risk Score with a sensitivity of 100.00% (95% CI: 94.87%, 100.00%) for identifying low risk chest pain patients and the Ultra-Low Risk Troponin Score which could be used in lieu of troponin. Future prospective studies will be required to externally validate these scores.

No description available.

<p>Deep learning plays an important role in various disciplines, such as auto-driving, information technology, manufacturing, medical studies, and financial studies. In the past decade, there have been fruitful studies on deep learning in which training and testing data are assumed to follow the same distribution to humans. Recent studies reveal that these dedicated models are vulnerable to adversarial attack, i.e., the predicting label may be changed even if the testing input has an unaware perturbation. However, most existing studies aim to develop computationally efficient adversarial learning algorithms without a thorough understanding of the statistical properties of these algorithms. This dissertation aims to provide theoretical understandings of adversarial training to figure out potential improvements in this area of research. </p> <p><br></p> <p>The first part of this dissertation focuses on the algorithmic stability of adversarial training. We reveal that the algorithmic stability of the vanilla adversarial training method is sub-optimal, and we study the effectiveness of a simple noise injection method. While noise injection improves stability, it also does not deteriorate the consistency of adversarial training.</p> <p><br></p> <p>The second part of this dissertation reveals a phase transition phenomenon in adversarial training. When the attack strength increases, the training trajectory of adversarial training will deviate from its natural counterpart. Consequently, various properties of adversarial training are different from clean training. It is essential to have adaptations in the training configuration and the neural network structure to improve adversarial training.</p> <p><br></p> <p>The last part of this dissertation focuses on how artificially generated data improves adversarial training. It is observed that utilizing synthetic data improves adversarial robustness, even if the data are generated using the original training data, i.e., no extra information is introduced. We use a theory to explain the reason behind this observation and propose further adaptations to utilize the generated data better.</p>

Malicious URL detection is important for cyber security experts and security agencies. With the drastic increase in internet usage, the distribution of such malware is a serious issue. Due to the wide variety of this malware, detection even with antivirus software is difficult. More than 12.8 million malicious URL websites are currently running. In this thesis, several machine learning classifiers along with ensemble methods are used to formulate a framework to detect this malware. Principal component analysis, k-fold cross-validation, and hyperparameter tuning are used to improve performance. A dataset from Kaggle is used for classification. Accuracy, precision, recall, and f-score are used as metrics to determine the model performance. Moreover, model behavior with a majority of one label in the dataset is also examined as is typical in the real world. / Graduate

<p> In this thesis we develop and evaluate automated "expert systems" for two applications: (i) gas/oil pipeline inspection using magnetic flux leakage information, (ii) treatment efficacy prediction and medical diagnosis using electroencephalograph (EEG) and clinical information. Both applications share the same methodology and procedure as they employ machine learning methods which learn their decision models using the training data (or past examples in real life/environment).</p> <p> The magnetic flux leakage (MFL) technique is commonly used for nondestructive testing (NDT) of oil and gas pipelines which are mostly buried underground. This testing involves the detection of metal defects and anomalies in the pipe wall, and the evaluation of the severity of these defects. The difficulty with the MFL method is the extent and complexity of the analysis of the MFL images. In this thesis we show how modern machine learning techniques can be used to considerable advantage in this respect.</p> <p> The problem of identifying in advance the most effective treatment agent for various psychiatric conditions remains an elusive goal. To address this challenge, an automated medical expert system is designed and then evaluated. The system is capable of predicting the treatment response for each individual patient at the outset of a therapy (i.e., using pre-treatment information) thus improving therapeutic efficiency and reducing personal and economic costs. Our experiments are focused on treatment planning and diagnosis of mood disorders and psychiatric illnesses. Through different experiments, we have shown that it is possible to predict treatment efficacy of a 'selective serotonin reuptake inhibitor' (SSRI) antidepressant and 'repetitive transcranial magnetic stimulation' (rTMS) therapies for patients with treatment-resistant major depressive disorder (MDD) or major depression. The predictions are based on pre-treatment quantitative EEG measurements. Also, prediction of post-treatment schizophrenia symptomatic scores, using pre-treatment EEG data, showed significant performance in patients treated with the drug clozapine. Clozapine is an antipsychotic medication of superior effectiveness in treating Schizophrenia but has several potentially severe side effects.</p> <p> Medical diagnosis is the second problem we consider in the neuroscience aspects of this thesis. In this research, an automated digital medical diagnosis methodology is developed to estimate/detect the type of a disease or illness that a patient is suffering. This intelligent diagnostic system can assist the physician/clinician by offering a second opinion on diagnosis. Several complex psychiatric illnesses may have many common symptoms and accurate diagnosis can, at times, be very difficult. Efficient diagnosis helps by avoiding prescription of wrong therapy /treatment to a patient. In our limited experiments, EEG data is used to make a diagnosis for distinguishing between various psychiatric illnesses including MDD, schizophrenia, and the depressed phase of bipolar affective disorder (BAD).</p> <p> In all problems considered in this thesis, specifically the neuroscience problem, a large number of candidate features are extracted from measurement data but most candidate features are found to be irrelevant and have little or no discriminative power. Finding a few most discriminating features that guarantee numerical efficiency and obtain a smooth and generalizable decision function, is a major challenge in this research. In this thesis, feature selection methods based on mutual information or Kullback-Leibler (KL) distance is employed to find the most statistically relevant features. For the multi-class diagnosis problem, to improve performance, a feature selection procedure denoted as feature combination feature selection is used which first finds discriminating features in all binary classification combinations, and then combines them into a larger feature subset to make a final multi-class decision. The two-dimensional (2D) representation of the feature data is also found to be useful for clustering analysis. The overall method was evaluated using a nested cross-validation procedure for which over 80% average prediction performance is obtained in all experiments. The results indicate that machine learning methods hold considerable promise in solving the challenging problems encountered in the two applications of concern.</p> / Thesis / Doctor of Philosophy (PhD)

Event Related Potential (ERP) measures derived from the electroencephalogram (EEG) have been widely used in research on language, cognition, and pathology. The high dimensionality (time x channel x condition) of a typical EEG/ERP dataset makes it a time-consuming prospect to properly analyze, explore, and validate knowledge without a particular restricted hypothesis. This study proposes an automated empirical greedy approach to the analysis process to datamine an EEG dataset for the location, robustness, and latency of ERPs, if any, present in a given dataset. We utilize Support Vector Machines (SVM), a well established machine learning model, on top of a preprocessing pipeline that focuses on detecting differences across experimental conditions. A hybrid of monte-carlo bootstrapping, cross-validation, and permutation tests is used to ensure the reproducibility of results. This framework serves to reduce researcher bias, time spent during analysis, and provide statistically sound results that are agnostic to dataset specifications including the ERPs in question. This method has been tested and validated on three different datasets with different ERPs (N100, Mismatch Negativity (MMN), N2b, Phonological Mapping Negativity (PMN), and P300). Results show statistically significant, above-chance level identification of all ERPs in their respective experimental conditions, latency, and location. / Thesis / Master of Science (MSc)