京都大学 / 新制・課程博士 / 博士(人間健康科学) / 甲第23825号 / 人健博第96号 / 新制||人健||7(附属図書館) / 京都大学大学院医学研究科人間健康科学系専攻 / (主査)教授 椎名 毅, 教授 精山 明敏, 教授 中本 裕士 / 学位規則第4条第1項該当 / Doctor of Human Health Sciences / Kyoto University / DFAM

Chess is a deep and intricate game, the master of which depends on learning tens of thousands of the patterns that may occur on the board. At Noctie, their mission is to aid this learning process through humanlike chess AI. A prominent challenge lies in curating instructive chess positions for students. Usually these are either manually found by going through large numbers of real games, or handcrafted – a time-consuming process. For effective learning, it is often useful to collect many positions following the same theme, or exhibiting the same type of pattern. Curating such collections from real games is a challenging task. This thesis investigates the transformer decoder-only architecture and its capability of generating realistic, game-like chess-positions. This investigation involved the development and training of a decoder model using Pytorch, and a simple web-based Turing test gaining larger understanding of testers experience. The developed chess model successfully generates chess positions, with constraining possibilities of fixed pieces, score intervals, and fixed empty positions. Controlled re-generation ensures satisfaction of score intervals, while empty positions are handled by iterating over the model's probabilities. Based on the limited data provided by the Turing test, the model seems to fool players below 2000 rank-points on chess.com, where guess percentages land near the 50 percent mark, providing no clear indication that it deviates from randomness.

<p dir="ltr">Modern machine learning research has recently made impressive progress across various research disciplines, such as computer vision, natural language processing, also in scientific fields including materials and molecule discovery, chip, and circuit design. In photonics/optics area, conventional methods in designing and optimiza- tion typically demand substantial time and extensive computing resources, where machine learning approaches hold the potential to significantly elevate and expe- dite these processes. On the other hand, machine learning algorithms can benefit from optical/photonics based neuromorphic computing systems due to their unique strengths in power consumption and parallelization. This talk will focus on imple- menting machine learning algorithms to optimize the optical/ photonics device (ML for photonics) as well as building optical based computing system for ML applica- tions (photonics for ML): First, I will discuss my work using probabilistic generative model (CVAE) for designing nanopatterned photonics power splitter with arbitrage splitting ratio. The model is incorporated with adversarial censoring and active learn- ing to increase the quality of generated devices. Next, I will report a physics-guided and physics-explainable recurrent neural network for time dynamics discovery in op- tical resonances, which can precisely forecast the time-domain response of resonance features with a very short portion of the initial input. The model is trained in a two-step multi-fidelity framework for high-accuracy forecast. In the end, I will present our progress in developing free space reconfigurable optical computing sys- tems for scientific computing, which is an optical based general matrix multiplication (GEMM) hardware accelerator by engineering a spatially reconfigurable array made from chalcogenide phase change materials. A device-system co-design methodology was implemented for GEMM system optimization. The device has been demonstrated over a various of ML applications.</p>

In modern society, cameras on intelligent devices can generate a huge amount of natural images, including images of the human body and face. Therefore, there is a huge social demand for more efficient editing of images to meet human production and life needs, including entertainment, such as image beauty. In recent years, Generative Models with Deep Learning techniques have attracted lots of attention in the Artificial Intelligence field, and some powerful methods, such as Variational Autoencoder and Generative Adversarial Networks, can generate very high-resolution and realistic images, especially for facial images, human body image. In this thesis, we follow the powerful generative model to achieve image generation and editing tasks, and we focus on human image generation and editing tasks, including local eye and face generation and editing, global human body generation, and editing. We introduce different methods to improve previous baselines based on different human regions. 1) Eye region of human image: Gaze correction and redirection aim to manipulate the eye gaze to a desired direction. Previous common gaze correction methods require annotating training data with precise gaze and head pose information. To address this issue, we proposed the new datasets as training data and formulated the gaze correction task as a generative inpainting problem, addressed using two new modules. 2) Face region of human image: Based on a powerful generative model for face region, many papers have learned to control the latent space to manipulate face attributes. However, they need more precise controls on 3d factors such as camera pose because they tend to ignore the underlying 3D scene rendering process. Thus, we take the pre-trained 3D-Aware generative model as the backbone and learn to manipulate the latent space using the attribute labels as conditional information to achieve the 3D-Aware face generation and editing task. 3) Human Body region of human image: 3D-Aware generative models have been shown to produce realistic images representing rigid/semi-rigid objects, such as facial regions. However, they usually struggle to generate high-quality images representing non-rigid objects, such as the human body, which greatly interests many computer graphics applications. Thus, we introduce semantic segmentation into the model. We split the entire generation pipeline into two stages and use intermediate segmentation masks to bridge these two stages. Furthermore, our model can control pose, semantic, and appearance codes by using multiple latent codes to achieve human image editing.

No description available.

Understanding how users interact with a company's service is essential for data-driven businesses that want to better cater to their users and improve their offering. By using a generative machine learning approach it is possible to model user behaviour and generate new data to simulate or recognize and explain typical usage patterns. In this work we introduce an approach for modelling users' interaction behaviour at scale in a client-service model. We propose a novel representation of multivariate time-series data as time pictures that express temporal correlations through spatial organization. This representation shares two key properties that convolutional networks have been built to exploit and allows us to develop an approach based on deep generative models that use convolutional networks as backbone. In introducing this approach of feature learning for time-series data, we expand the application of convolutional neural networks in the multivariate time-series domain, and specifically user interaction data. We adopt a variational approach inspired by the β-VAE framework in order to learn hidden factors that define different user behaviour patterns. We explore different values for the regularization parameter β and show that it is possible to construct a model that learns a latent representation of identifiable and different user behaviours. We show on real-world data that the model generates realistic samples, that capture the true population-level statistics of the interaction behaviour data, learns different user behaviours, and provides accurate imputations of missing data. / Förståelse för hur användare interagerar med ett företags tjänst är essentiell för data-drivna affärsverksamheter med ambitioner om att bättre tillgodose dess användare och att förbättra deras utbud. Generativ maskininlärning möjliggör modellering av användarbeteende och genererande av ny data i syfte att simulera eller identifiera och förklara typiska användarmönster. I detta arbete introducerar vi ett tillvägagångssätt för storskalig modellering av användarinteraktion i en klientservice-modell. Vi föreslår en ny representation av multivariat tidsseriedata i form av tidsbilder vilka representerar temporala korrelationer via spatial organisering. Denna representation delar två nyckelegenskaper som faltningsnätverk har utvecklats för att exploatera, vilket tillåter oss att utveckla ett tillvägagångssätt baserat på på djupa generativa modeller som bygger på faltningsnätverk. Genom att introducera detta tillvägagångssätt för tidsseriedata expanderar vi applicering av faltningsnätverk inom domänen för multivariat tidsserie, specifikt för användarinteraktionsdata. Vi använder ett tillvägagångssätt inspirerat av ramverket β-VAE i syfte att lära modellen gömda faktorer som definierar olika användarmönster. Vi utforskar olika värden för regulariseringsparametern β och visar att det är möjligt att konstruera en modell som lär sig en latent representation av identifierbara och multipla användarbeteenden. Vi visar med verklig data att modellen genererar realistiska exempel vilka i sin tur fångar statistiken på populationsnivå hos användarinteraktionsdatan, samt lär olika användarbeteenden och bidrar med precisa imputationer av saknad data.

The Automatic Dependent Surveillance-Broadcast (ADS-B) is a key component of the Next Generation Air Transportation System (Next Gen) that manages the increasingly congested airspace and operation. From Jan 2020, the U.S. Federal Aviation Administration (FAA) mandated the use of (ADS-B) as a key component of Next Gen project. ADS-Bprovides accurate aircraft localization via satellite navigation and efficient air traffic management, and also improves the safety of thousands of passengers travelling worldwide. While the benefits of ADS-B are well known, the fact that ADS-B is an open protocol introduces various exploitable security vulnerabilities. One practical threat is the ADS-B spoofing attack that targets the ground station, in which the ground-based attacker manipulates the International Civil Aviation Organization (ICAO) address (which is a unique identifierfor each aircraft) in the ADS-B forwarded messages to fake the appearance of non-existent aircraft or masquerade as a trusted aircraft. As a result, this type of attack can confuseand misguide the aircraft pilots or the air traffic control personnel and cause dangerous maneuvers. In this project, we intend to build a robust Intrusion Detection System (IDS) to detectanomalous behavior and classify attacks in an aircraft ADS-B protocol in real time duringair-ground communication. The IDS system we propose is a 3 stage deep learning framework built using Spatial Graph Convolution Networks and Deep auto-regressive generative model. In stage 1 we use a Graph convolution network architecture to classify the dataas attacked or normal in the entire airspace of an operating aircraft. In stage 2 we analyze the sequences of air-space states to identify anomalies using a generative Wavenet modeland simultaneously output feature under attack. Final stage consist of aircraft (ICAO) classification module based on unique RF transmitter signal characteristics of an aircraft. This allows the ground station operator to examine each incoming message based on the Phylayer features as well as message data field (such as, position, velocity, altitude) and flagsuspicious messages. The model is trained in a supervised fashion using federated learning where the data remains private to the data owner, i.e.: aircraft-ground station without data being explicitly sent to the cloud server. The server only receives the learned parameters for inference, there by training the entire model on the edge, thus preserving data-privacyand potential adversarial attacks. We aim to achieve a high precision real-time IDS system, with very low false alarm rate for real world deployment

Pas de résumé en français / It is generally maintained that one of cortex’ functions is the storage of a large number of memories. In this picture, the physical substrate of memories is thought to be realised in pattern and strengths of synaptic connections among cortical neurons. Memory recall is associated with neuronal activity that is shaped by this connectivity. In this framework, active memories are represented by attractors in the space of neural activity. Electrical activity in cortical neurones in vivo exhibits prominent temporal irregularity. A standard way to account for this phenomenon is to postulate that recurrent synaptic excitation and inhibition as well as external inputs are balanced. In the common view, however, these balanced networks do not easily support the coexistence of multiple attractors. This is problematic in view of memory function. Recently, theoretical studies showed that balanced networks with synapses that exhibit short-term plasticity (STP) are able to maintain multiple stable states. In order to investigate whether experimentally obtained synaptic parameters are consistent with model predictions, we developed a new methodology that is capable to quantify both response variability and STP at the same synapse in an integrated and statistically-principled way. This approach yields higher parameter precision than standard procedures and allows for the use of more efficient stimulation protocols. However, the findings with respect to STP parameters do not allow to make conclusive statements about the validity of synaptic theories of balanced working memory. In the second part of this thesis an alternative theory of cortical memory storage is developed. The theory is based on the assumptions that memories are stored in attractor networks, and that memories are not represented by network states differing in their average activity levels, but by micro-states sharing the same global statistics. Different memories differ with respect to their spatial distributions of firing rates. From this the main result is derived: the balanced state is a necessary condition for extensive memory storage. Furthermore, we analytically calculate memory storage capacities of rate neurone networks. Remarkably, it can be shown that crucial properties of neuronal activity and physiology that are consistent with experimental observations are directly predicted by the theory if optimal memory storage capacity is required.

L’apprentissage automatique et spécialement l’apprentissage profond se sont imposés ces dernières années pour résoudre une large variété de tâches. Une des applications les plus remarquables concerne la vision par ordinateur. Les systèmes de détection ou de classification ont connu des avancées majeurs grâce a l’apprentissage profond. Cependant, il reste de nombreux obstacles à une compréhension du monde similaire aux être vivants. Ces derniers n’ont pas besoin de labels pour classifier, pour extraire des caractéristiques du monde réel. L’apprentissage non supervisé est un des axes de recherche qui se concentre sur la résolution de ce problème. Dans ce mémoire, je présente un nouveau moyen d’entrainer des réseaux de neurones de manière non supervisée. Je présente une méthode permettant d’échantillonner de manière itérative a partir de bruit afin de générer des données qui se rapprochent des données d’entrainement. Cette procédure itérative s’appelle l’entrainement par infusion qui est une nouvelle approche permettant d’apprendre l’opérateur de transition d’une chaine de Markov. Dans le premier chapitre, j’introduis des bases concernant l’apprentissage automatique et la théorie des probabilités. Dans le second chapitre, j’expose les modèles génératifs qui ont inspiré ce travail. Dans le troisième et dernier chapitre, je présente comment améliorer l’échantillonnage dans les modèles génératifs avec l’entrainement par infusion. / Machine learning and specifically deep learning has made significant breakthroughs in recent years concerning different tasks. One well known application of deep learning is computer vision. Tasks such as detection or classification are nearly considered solved by the community. However, training state-of-the-art models for such tasks requires to have labels associated to the data we want to classify. A more general goal is, similarly to animal brains, to be able to design algorithms that can extract meaningful features from data that aren’t labeled. Unsupervised learning is one of the axes that try to solve this problem. In this thesis, I present a new way to train a neural network as a generative model capable of generating quality samples (a task akin to imagining). I explain how by starting from noise, it is possible to get samples which are close to the training data. This iterative procedure is called Infusion training and is a novel approach to learning the transition operator of a generative Markov chain. In the first chapter, I present some background about machine learning and probabilistic models. The second chapter presents generative models that inspired this work. The third and last chapter presents and investigates our novel approach to learn a generative model with Infusion training.

The brain is a complex organ that is computationally strong. Recent research in the field of neurobiology help scientists to better understand the working of the brain, especially how the brain represents or codes external signals. The research shows that the neural code is sparse. A sparse code is a code in which few neurons participate in the representation of a signal. Neurons communicate with each other by sending pulses or spikes at certain times. The spikes send between several neurons over time is called a spike train. A spike train contains all the important information about the signal that it codes. This thesis shows how sparse coding can be used to do speech recognition. The recognition process consists of three parts. First the speech signal is transformed into a spectrogram. Thereafter a sparse code to represent the spectrogram is found. The spectrogram serves as the input to a linear generative model. The output of themodel is a sparse code that can be interpreted as a spike train. Lastly a spike train model recognises the words that are encoded in the spike train. The algorithms that search for sparse codes to represent signals require many computations. We therefore propose an algorithm that is more efficient than current algorithms. The algorithm makes it possible to find sparse codes in reasonable time if the spectrogram is fairly coarse. The system achieves a word error rate of 19% with a coarse spectrogram, while a system based on Hidden Markov Models achieves a word error rate of 15% on the same spectrograms. / Thesis (PhD)--University of Pretoria, 2008. / Electrical, Electronic and Computer Engineering / unrestricted