No description available.

京都大学 / 新制・課程博士 / 博士(情報学) / 甲第24729号 / 情博第817号 / 新制||情||137(附属図書館) / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 河原 達也, 教授 黒橋 禎夫, 教授 森 信介 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Multi-Task Learning (MTL) has been an active research area in machine learning for two decades. By training multiple relevant tasks simultaneously with information shared across tasks, it is possible to improve the generalization performance of each task, compared to training each individual task independently. During the past decade, most MTL research has been based on the Regularization-Loss framework due to its flexibility in specifying various types of information sharing strategies, the opportunity it offers to yield a kernel-based methods and its capability in promoting sparse feature representations. However, certain limitations exist in both theoretical and practical aspects of Regularization-Loss-based MTL. Theoretically, previous research on generalization bounds in connection to MTL Hypothesis Space (HS)s, where data of all tasks are pre-processed by a (partially) common operator, has been limited in two aspects: First, all previous works assumed linearity of the operator, therefore completely excluding kernel-based MTL HSs, for which the operator is potentially non-linear. Secondly, all previous works, rather unnecessarily, assumed that all the task weights to be constrained within norm-balls, whose radii are equal. The requirement of equal radii leads to significant inflexibility of the relevant HSs, which may cause the generalization performance of the corresponding MTL models to deteriorate. Practically, various algorithms have been developed for kernel-based MTL models, due to different characteristics of the formulations. Most of these algorithms are a burden to develop and end up being quite sophisticated, so that practitioners may face a hard task in interpreting and implementing them, especially when multiple models are involved. This is even more so, when Multi-Task Multiple Kernel Learning (MT-MKL) models are considered. This research largely resolves the above limitations. Theoretically, a pair of new kernel-based HSs are proposed: one for single-kernel MTL, and another one for MT-MKL. Unlike previous works, we allow each task weight to be constrained within a norm-ball, whose radius is learned during training. By deriving and analyzing the generalization bounds of these two HSs, we show that, indeed, such a flexibility leads to much tighter generalization bounds, which often results to significantly better generalization performance. Based on this observation, a pair of new models is developed, one for each case: single-kernel MTL, and another one for MT-MKL. From a practical perspective, we propose a general MT-MKL framework that covers most of the prominent MT-MKL approaches, including our new MT-MKL formulation. Then, a general purpose algorithm is developed to solve the framework, which can also be employed for training all other models subsumed by this framework. A series of experiments is conducted to assess the merits of the proposed mode when trained by the new algorithm. Certain properties of our HSs and formulations are demonstrated, and the advantage of our model in terms of classification accuracy is shown via these experiments.

The purpose of this research was to examine operator workload and performance in a high risk, multi-task environment. Specifically, the research examined if a gunner of a Future Combat System, such as a Mounted Combat System, could effectively detect targets in the immediate environment while concurrently operating robotic assets in a remote environment. It also analyzed possible effects of individual difference factors, such as spatial ability and attentional control, on operator performance and workload. The experimental conditions included a gunner baseline and concurrent task conditions where participants simultaneously performed gunnery tasks and one of the following tasks: monitor an unmanned ground vehicle (UGV) via a video feed (Monitor), manage a semi-autonomous UGV, and teleoperate a UGV (Teleop). The analysis showed that the asset condition significantly impacted gunnery performance with the gunner baseline having the highest number of targets detected (M = 13.600 , SD = 2.353), and concurrent Teleop condition the lowest (M = 9.325 , SD = 2.424). The research also found that high spatial ability participants tended to detect more targets than low spatial ability participants. Robotic task performance was also affect by the asset condition. The results showed that the robotic target detection rate was lower for the concurrent task conditions. A significant difference was seen between the UGV-baseline (80.1%) when participants performed UGV tasks only and UGV-concurrent conditions (67.5%) when the participants performed UGV tasks concurrently with gunnery tasks. Overall, this study revealed that there were performance decrements for the gunnery tasks as well as the robotic tasks when the tasks were performed concurrently.

This thesis focuses on integrating learning paradigms that ‘share to learn,’ i.e., Multitask Learning (MTL), and ‘learn (how) to share,’ i.e., meta learning. MTL involves learning several tasks simultaneously within a shared network structure so that the tasks can mutually benefit each other’s learning. While meta learning, better known as ‘learning to learn,’ is an approach to reducing the amount of time and computation required to learn a novel task by leveraging on knowledge accumulated over the course of numerous training episodes of various tasks. The learning process in the human brain is innate and natural. Even before birth, it is capable of learning and memorizing. As a consequence, humans do not learn everything from scratch, and because they are naturally capable of effortlessly transferring their knowledge between tasks, they quickly learn new skills. Humans naturally tend to believe that similar tasks have (somewhat) similar solutions or approaches, so sharing knowledge from a previous activity makes it feasible to learn a new task quickly in a few tries. For instance, the skills acquired while learning to ride a bike are helpful when learning to ride a motorbike, which is, in turn, helpful when learning to drive a car. This natural learning process, which involves sharing information between tasks, has inspired a few research areas in Deep Learning (DL), such as transfer learning, MTL, meta learning, Lifelong Learning (LL), and many more, to create similar neurally-weighted algorithms. These information-sharing algorithms exploit the knowledge gained from one task to improve the performance of another related task. However, they vary in terms of what information they share, when to share, and why to share. This thesis focuses particularly on MTL and meta learning, and presents a comprehensive explanation of both the learning paradigms. A theoretical comparison of both algorithms demonstrates that the strengths of one can outweigh the constraints of the other. Therefore, this work aims to combine MTL and meta learning to attain the best of both worlds. The main contribution of this thesis is Multi-task Meta Learning (MTML), an integration of MTL and meta learning. As the gradient (or optimization) based metalearning follows an episodic approach to train a network, we propose multi-task learning episodes to train a MTML network in this work. The basic idea is to train a multi-task model using bi-level meta-optimization so that when a new task is added, it can learn in fewer steps and perform at least as good as traditional single-task learning on the new task. The MTML paradigm is demonstrated on two publicly available datasets – the NYU-v2 and the taskonomy dataset, for which four tasks are considered, i.e., semantic segmentation, depth estimation, surface normal estimation, and edge detection. This work presents a comparative empirical analysis of MTML to single-task and multi-task learning, where it is evident that MTML excels for most tasks. The future direction of this work includes developing efficient and autonomous MTL architectures by exploiting the concepts of meta learning. The main goal will be to create a task-adaptive MTL, where meta learning may learn to select layers (or features) from the shared structure for every task because not all tasks require the same highlevel, fine-grained features from the shared network. This can be seen as another way of combining MTL and meta learning. It will also introduce modular learning in the multi-task architecture. Furthermore, this work can be extended to include multi-modal multi-task learning, which will help to study the contributions of each input modality to various tasks.

Many countries around the world are undergoing massive urbanization campaigns at an unprecedented rate, heralded by promises of economical prosperity and bolstered population health and well-being. Projections indicate that by 2050, nearly 68% of the world populace will reside in these urban environments. However, rapid growth at such an exceptional scale poses unique challenges pertaining to environmental quality and food production, which can negate the effectiveness of the aforementioned boons. As such, there is an emphasis on mitigating these negative effects through the construction of smart and connected communities (S&CC), which integrate both artificial intelligence (AI) and the Internet of Things (IoT). This coupling of intelligent technologies also poses interesting system design challenges pertaining to the fusion of the diverse, heterogeneous datasets available to IoT environments, and the ability to learn multiple S&CC problem sets concurrently. Attention-based Transformer networks are of particular interest given their success across diverse fields of natural language processing (NLP), computer vision, time-series regression, and multi-modal data fusion in recent years. This begs the question whether Transformers can be further diversified to leverage fusions of IoT data sources for heterogeneous multi-task learning in S&CC trade spaces. This is a fundamental question that this thesis seeks to answer. Indeed, the key contribution of this thesis is the design and application of Transformer networks for developing AI systems in emerging smart cities. This is executed within a collaborative U.S.-Indonesia effort between Virginia Tech, the University of Colorado Boulder, the Universitas Gadjah Mada, and the Institut Teknologi Bandung with the goal of growing smart and sustainable garden alleys in Makassar City, Indonesia. Specifically, a proof-of-concept AI nerve-center is proposed using a backbone of pure-encoder Transformer architectures to learn a diverse set of tasks such as multivariate time-series regression, visual plant disease classification, and image-time-series fusion. To facilitate the data fusion tasks, an effective algorithm is also proposed to synthesize heterogeneous feature sets, such as multivariate time-series and time-correlated images. Moreover, a hyperparameter tuning framework is also proposed to standardize and automate model training regimes. Extensive experimentation shows that the proposed Transformer-based systems can handle various input data types via custom sequence embedding techniques, and are naturally suited to learning a diverse set of tasks. Further, the results also show that multi-task learners increase both memory and computational efficiency while maintaining comparable performance to both single-task variants, and non-Transformer baselines. This demonstrates the flexibility of Transformer networks to learn from a fusion of IoT data sources, their applicability in S&CC trade spaces, and their further potential for deployment on edge computing devices. / Master of Science / Many countries around the world are undergoing massive urbanization campaigns at an unprecedented rate, heralded by promises of economical prosperity and bolstered population health and well-being. Projections indicate that by 2050, nearly 68% of the world populace will reside in these urban environments. However, rapid growth at such an exceptional scale poses unique environmental and food cultivation challenges. Hence, there is a focus on reducing these negative effects through building smart and connected communities (S&CC). The term connected is derived from the integration of small, low-cost devices which gather information from the surrounding environment, called the Internet of Things (IoT). Likewise, smart is a term derived from the integration of artificial intelligence (AI), which is used to make informed decisions based on IoT-collected information. This coupling of intelligent technologies also poses its own unique challenges pertaining to the blending of IoT data with highly diverse characteristics. Of specific interest is the design of AI models that can not only learn from a fusion of this diverse information, but also learn to perform multiple tasks in parallel. Attention-based networks are a relatively new category of AI which learn to focus on, or attend to, the most important portions of an arbitrary data sequence. Transformers are AI models which are designed using attention as their backbone, and have been employed to much success in many fields in recent years. This success begs the question whether Transformers can be further extended to put the smart in S&CC. The overarching goal of this thesis is to design and implement a Transformer-based AI system for emerging smart cities. In particular, this is accomplished within a U.S.-Indonesia collaborative effort with the goal of growing smart and sustainable garden alleys in Makassar City, Indonesia.

Generalized collaborative drones are a technology that has many potential benefits. General purpose drones that can handle exploration, navigation, manipulation, and more without having to be reprogrammed would be an immense breakthrough for usability and adoption of the technology. The ability to develop these multi-task, multi-agent drone systems is limited by the lack of available training environments, as well as deficiencies of multi-task learning due to a phenomenon known as catastrophic forgetting. In this thesis, we present a set of simulation environments for exploring the abilities of multi-task drone systems and provide a platform for testing agents in incremental single-agent and multi-agent learning scenarios. The multi-task platform is an extension of an existing drone simulation environment written in Python using the PyBullet Physics Simulation Engine, with these environments incorporated. Using this platform, we present an analysis of Incremental Learning and detail the beneficial impacts of using the technique for multi-task learning, with respect to multi-task learning speed and catastrophic forgetting. Finally, we introduce a novel algorithm, Incremental Learning with Second-Order Approximation Regularization (IL-SOAR), to mitigate some of the effects of catastrophic forgetting in multi-task learning. We show the impact of this method and contrast the performance relative to a multi-agent multi-task approach using a centralized policy sharing algorithm. / Master of Science / Machine Learning techniques allow drones to be trained to achieve tasks which are otherwise time-consuming or difficult. The goal of this thesis is to facilitate the work of creating these complex drone machine learning systems by exploring Reinforcement Learning (RL), a field of machine learning which involves learning the correct actions to take through experience. Currently, RL methods are effective in the design of drones which are able to solve one particular task. The next step in this technology is to develop RL systems which are able to handle generalization and perform well across multiple tasks. In this thesis, simulation environments for drones to learn complex tasks are created, and algorithms which are able to train drones in multiple hard tasks are developed and tested. We explore the benefits of using a specific multi-task training technique known as Incremental Learning. Additionally, we consider one of the prohibitive factors of multi-task machine learning-based solutions, the degradation problem of agent performance on previously learned tasks, known as catastrophic forgetting. We create an algorithm that aims to prevent the impact of forgetting when training drones sequentially on new tasks. We contrast this approach with a multi-agent solution, where multiple drones learn simultaneously across the tasks.

Creating embodied agents capable of performing a wide range of tasks in different types of environments has been a longstanding challenge in deep reinforcement learning. A novel network architecture introduced in 2021 called the Active Dendrite Network [A. Iyer et al., “Avoiding Catastrophe: Active Dendrites Enable Multi-Task Learning in Dynamic Environments”] designed to create sparse subnetworks for different tasks showed promising multi-tasking performance on the Meta-World [T. Yu et al., “Meta-World: A Benchmark and Evaluation for Multi-Task and Meta Reinforcement Learning”] multi-tasking benchmark. This thesis further explores the performance of this novel architecture in a multi-tasking environment focused on physical animations and locomotion. Specifically we implement and compare the architecture to the commonly used Multi-Layer Perceptron (MLP) architecture on a multi-task reinforcement learning problem in a video-game setting consisting of training a hexapedal agent on a set of locomotion tasks involving moving at different speeds, turning and standing still. The evaluation focused on two areas: (1) Assessing the average overall performance of the Active Dendrite Network relative to the MLP on a set of locomotive scenarios featuring our behaviour sets and environments. (2) Assessing the relative impact Active Dendrite networks have on transfer learning between related tasks by comparing their performance on novel behaviours shortly after training a related behaviour. Our findings suggest that the novel Active Dendrite Network can make better use of limited network capacity compared to the MLP - the Active Dendrite Network outperformed the MLP by ∼18% on our benchmark using limited network capacity. When both networks have sufficient capacity however, there is not much difference between the two. We further find that Active Dendrite Networks have very similar transfer-learning capabilities compared to the MLP in our benchmarks.

The original intention of the recommender system is to solve the problem of information explosion, hoping to help users find the content they need more efficiently. In an e-commerce platform, users typically interact with items that they are interested in or need in a variety of ways. For example, buying, browsing details, etc. These interactions are recorded as time-series information. How to use this sequential information to predict user behaviors in the future and give an efficient and effective recommendation is a very important problem. For content providers, such as merchants in e-commerce platforms, more accurate recommendation means higher traffic, CTR (click-through rate), and revenue. Therefore, in the industry, the CTR model for recommendation systems is a research hotspot. However, in the fine ranking stage of the recommendation system, the existing models have some limitations. No researcher has attempted to predict multiple behaviors of one user simultaneously by processing sequential information. We define this problem as the multi-task sequential recommendation problem. In response to this problem, we study the CTR model, sequential recommendation, and multi-task learning. Based on these studies, this paper proposes AMBSN (Attention-based Multi-Behavior Sequential Network). Specifically, we added a transformer layer, the activation unit, and the multi-task tower to the traditional Embedding&MLP (multi-layer perceptron) model. The transformer layer enables our model to efficiently extract sequential behavior information, the activation unit can understand user interests, and the multi-task tower structure makes the model give the prediction of different user behaviors at the same time. We choose user behavior data from Taobao for recommendation published on TianChi as the dataset, and AUC as the evaluation criterion. We compare the performance of AMBSN and some other models on the test set after training. The final results of the experiment show that our model outperforms some existing models. / L’intenzione originale del sistema di raccomandazione è risolvere il problema dell’esplosione delle informazioni, sperando di aiutare gli utenti a trovare il contenuto di cui hanno bisogno in modo più efficiente. In una piattaforma di e-commerce, gli utenti in genere interagiscono con gli articoli a cui sono interessati o di cui hanno bisogno in vari modi. Ad esempio, acquisti, dettagli di navigazione, ecc. Queste interazioni vengono registrate come informazioni di serie temporali. Come utilizzare queste informazioni sequenziali per prevedere i comportamenti degli utenti in futuro e fornire una raccomandazione efficiente ed efficace è un problema molto importante. Per i fornitori di contenuti, come i commercianti nelle piattaforme di e-commerce, una raccomandazione più accurata significa traffico, CTR (percentuale di clic) ed entrate più elevati. Pertanto, nel settore, il modello CTR per i sistemi di raccomandazione è un hotspot di ricerca. Tuttavia, nella fase di classificazione fine del sistema di raccomandazione, i modelli esistenti presentano alcune limitazioni. Nessun ricercatore ha tentato di prevedere più comportamenti di un utente contemporaneamente elaborando informazioni sequenziali. Definiamo questo problema come il problema di raccomandazione sequenziale multi-task. In risposta a questo problema, studiamo il modello CTR, la raccomandazione sequenziale e l’apprendimento multi-task. Sulla base di questi studi, questo documento propone AMBSN (Attention-based Multi-Behavior Sequential Network). In particolare, abbiamo aggiunto uno strato trasformatore, l’unità di attivazione e la torre multi-task al tradizionale modello Embedding&MLP (multi-layer perceptron). Il livello del trasformatore consente al nostro modello di estrarre in modo efficiente le informazioni sul comportamento sequenziale, l’unità di attivazione può comprendere gli interessi degli utenti e la struttura della torre multi-task fa sì che il modello fornisca la previsione di diversi comportamenti degli utenti contemporaneamente. Scegliamo i dati sul comportamento degli utenti da Taobao per la raccomandazione pubblicata su TianChi come set di dati e l’AUC come criterio di valutazione. Confrontiamo le prestazioni di AMBSN e di alcuni altri modelli sul set di test dopo l’allenamento. I risultati finali dell’esperimento mostrano che il nostro modello supera alcuni modelli esistenti.

The perception tasks in autonomous driving, namely 3D object detection and map segmentation, play a crucial role in enabling vehicles to perceive the surrounding environment. The traditional approach is to have a single network for each task and complete all tasks in a sequential manner. However, this method suffers from repeated feature extraction and error propagation, leading to inefficiency and reduced accuracy. While multi-task learning can eliminate redundant computations and facilitate information exchange among tasks, improving efficiency and overall system performance, it can also lead to a reduction in the performance of a particular task, compared to single-task training due to gradient dominance. To tackle this problem, this thesis aims to bridge the performance gap between multi-task and single-task learning. We first utilize the GradNorm method to dynamically readjust the loss weights while training. We further add individual encoders to allow fine-grained feature learning for each task. Based on the existing perception network, we adapt our dynamic loss strategy and new encoder architecture, which shows that our results match or even surpass the performance of each task in a multitask setting, compared to the single task. We also evaluate the computational efficiency of our method, further demonstrating the advantages of multi-task learning in the autonomous driving domain where real-time computing is non-negotiable. / Perception inom autonom körning, det vill säga 3D objektsdetektion och map segmentation, spelar en avgörande roll för att möjliggöra att fordon uppfattar den omgivande miljön. Traditionellt sett utförs detta av en grupp nätverk, en för varje uppgift, där uppgifterna utförs sekventiellt. Detta skapar problem i form av upprepad extrahering av egenskaper i datan och felfortplantning vilket leder till försämrad beräkningshastighet och resultat. Multi-task learning kan eliminera onödiga beräkningar och möjliggöra utbyte av information mellan uppgifter, vilket medför förbättringar inom effektivitet och systemets generella prestanda i relation till single-task learning. Däremot kan det leda till försämrade resultat i enskilda uppgifter på grund av gradient dominans. Denna avhandling syftar till att bemöta detta problem genom att överbrygga gapet mellan multi-task och single-task learning. Vi använder oss av GradNorm metoden för att dynamiskt justera gradienternas magnitud under träning. Enskilda encoders lades till i vardera nätverk för att möjliggöra fine-grained feature learning för varje uppgift. GradNorm och ytterligare encoders applicerades på det befintliga perception-nätverket, vilket gav resultat som är lika bra eller bättre vid multi-task learning som single-task learning för varje uppgift. Även effektivitet vid beräkning utvärderades, vilket ytterligare visade fördelarna av att använda multi-task learning i autonom körning där beräkning i realtid är av högsta prioritet.