One of the main problems in machine learning is to train a predictive model from training data and to make predictions on test data. Most predictive models are constructed under the assumption that the training data is governed by the exact same distribution which the model will later be exposed to. In practice, control over the data collection process is often imperfect. A typical scenario is when labels are collected by questionnaires and one does not have access to the test population. For example, parts of the test population are underrepresented in the survey, out of reach, or do not return the questionnaire. In many applications training data from the test distribution are scarce because they are difficult to obtain or very expensive. Data from auxiliary sources drawn from similar distributions are often cheaply available. This thesis centers around learning under differing training and test distributions and covers several problem settings with different assumptions on the relationship between training and test distributions-including multi-task learning and learning under covariate shift and sample selection bias. Several new models are derived that directly characterize the divergence between training and test distributions, without the intermediate step of estimating training and test distributions separately. The integral part of these models are rescaling weights that match the rescaled or resampled training distribution to the test distribution. Integrated models are studied where only one optimization problem needs to be solved for learning under differing distributions. With a two-step approximation to the integrated models almost any supervised learning algorithm can be adopted to biased training data. In case studies on spam filtering, HIV therapy screening, targeted advertising, and other applications the performance of the new models is compared to state-of-the-art reference methods. / Eines der wichtigsten Probleme im Maschinellen Lernen ist das Trainieren von Vorhersagemodellen aus Trainingsdaten und das Ableiten von Vorhersagen für Testdaten. Vorhersagemodelle basieren üblicherweise auf der Annahme, dass Trainingsdaten aus der gleichen Verteilung gezogen werden wie Testdaten. In der Praxis ist diese Annahme oft nicht erfüllt, zum Beispiel, wenn Trainingsdaten durch Fragebögen gesammelt werden. Hier steht meist nur eine verzerrte Zielpopulation zur Verfügung, denn Teile der Population können unterrepräsentiert sein, nicht erreichbar sein, oder ignorieren die Aufforderung zum Ausfüllen des Fragebogens. In vielen Anwendungen stehen nur sehr wenige Trainingsdaten aus der Testverteilung zur Verfügung, weil solche Daten teuer oder aufwändig zu sammeln sind. Daten aus alternativen Quellen, die aus ähnlichen Verteilungen gezogen werden, sind oft viel einfacher und günstiger zu beschaffen. Die vorliegende Arbeit beschäftigt sich mit dem Lernen von Vorhersagemodellen aus Trainingsdaten, deren Verteilung sich von der Testverteilung unterscheidet. Es werden verschiedene Problemstellungen behandelt, die von unterschiedlichen Annahmen über die Beziehung zwischen Trainings- und Testverteilung ausgehen. Darunter fallen auch Multi-Task-Lernen und Lernen unter Covariate Shift und Sample Selection Bias. Es werden mehrere neue Modelle hergeleitet, die direkt den Unterschied zwischen Trainings- und Testverteilung charakterisieren, ohne dass eine einzelne Schätzung der Verteilungen nötig ist. Zentrale Bestandteile der Modelle sind Gewichtungsfaktoren, mit denen die Trainingsverteilung durch Umgewichtung auf die Testverteilung abgebildet wird. Es werden kombinierte Modelle zum Lernen mit verschiedenen Trainings- und Testverteilungen untersucht, für deren Schätzung nur ein einziges Optimierungsproblem gelöst werden muss. Die kombinierten Modelle können mit zwei Optimierungsschritten approximiert werden und dadurch kann fast jedes gängige Vorhersagemodell so erweitert werden, dass verzerrte Trainingsverteilungen korrigiert werden. In Fallstudien zu Email-Spam-Filterung, HIV-Therapieempfehlung, Zielgruppenmarketing und anderen Anwendungen werden die neuen Modelle mit Referenzmethoden verglichen.

This thesis explores multi-task learning for combustion flame characterization i.e to learn different characteristics of the combustion flame. We propose a multi-task convolutional neural network for two tasks i.e. PFR (Pilot fuel ratio) and fuel type classification based on the images of stable combustion. We utilize transfer learning and adopt VGG16 to develop a multi-task convolutional neural network to jointly learn the aforementioned tasks. We also compare the performance of the individual CNN model for two tasks with multi-task CNN which learns these two tasks jointly by sharing visual knowledge among the tasks. We share the effectiveness of our proposed approach to a private company’s dataset. To the best of our knowledge, this is the first work being done for jointly learning different characteristics of the combustion flame. / <p>This wrok as done with Siemens, and we have applied for a patent which is still pending.</p>

The Random Forests ensemble predictor has proven to be well-suited for solving a multitudeof different prediction problems. In this thesis, we propose an extension to the Random Forestframework that allows Random Forests to be constructed for multi-output decision problemswith arbitrary combinations of classification and regression responses, with the goal ofincreasing predictive performance for such multi-output problems. We show that our methodfor combining decision tasks within the same decision tree reduces prediction error for mosttasks compared to single-output decision trees based on the same node impurity metrics, andprovide a comparison of different methods for combining such metrics. / Program: Magisterutbildning i informatik

In the context of control, the motion of a legged robot is very challenging compared with traditional fixed manipulator. Recently, many researches have been conducted to control the motion of legged robot with different techniques. On the other hand, manipulation tasks have been addressed in many applications. These researches solved either the mobility or the manipulation problems, but integrating both properties in one system is still not available. In this thesis, a control algorithm is presented to control both locomotion and manipulation in a six legged robot. Landmines detection process is considered as a case study of this project to accelerate the mine detection operation by performing both walking and scanning simultaneously. In order to qualify the robot to perform more tasks in addition to the walking task, the joint redundancy of the robot is exploited optimally. The tasks are arranged according to their importance to high level of priority and low level of priority. A new task priority redundancy resolution technique is developed to overcome the effect of the algorithmic singularities and the kinematic singularity. The computational aspects of the solution are also considered in view of a real-time implementation. Due to the dynamic changes in the size of the robot motion space, the algorithm has the ability to make a trade-off between the number of achieved tasks and the imposed constraints. Furthermore, an appropriate hierarchy is imposed in order to ensure an accurate decoupling between the executed tasks. The dynamic effect of the arm on the overall performance of the robot is attenuated by reducing the optimisation variables. The effectiveness of the method is evaluated on a Computer Aided Design (CAD) model and the simulations of the whole operation are conducted using MATLAB and SimMechanics.

<p>In most machine learning approaches, it is usually assumed that data are complete. When data are partially missing due to various reasons, for example, the failure of a subset of sensors, image corruption or inadequate medical measurements, many learning methods designed for complete data cannot be directly applied. In this dissertation we treat two kinds of problems with incomplete data using non-parametric Bayesian approaches: classification with incomplete features and analysis of low-rank matrices with missing entries.</p><p>Incomplete data in classification problems are handled by assuming input features to be generated from a mixture-of-experts model, with each individual expert (classifier) defined by a local Gaussian in feature space. With a linear classifier associated with each Gaussian component, nonlinear classification boundaries are achievable without the introduction of kernels. Within the proposed model, the number of components is theoretically ``infinite'' as defined by a Dirichlet process construction, with the actual number of mixture components (experts) needed inferred based upon the data under test. With a higher-level DP we further extend the classifier for analysis of multiple related tasks (multi-task learning), where model components may be shared across tasks. Available data could be augmented by this way of information transfer even when tasks are only similar in some local regions of feature space, which is particularly critical for cases with scarce incomplete training samples from each task. The proposed algorithms are implemented using efficient variational Bayesian inference and robust performance is demonstrated on synthetic data, benchmark data sets, and real data with natural missing values.</p><p>Another scenario of interest is to complete a data matrix with entries missing. The recovery of missing matrix entries is not possible without additional assumptions on the matrix under test, and here we employ the common assumption that the matrix is low-rank. Unlike methods with a preset fixed rank, we propose a non-parametric Bayesian alternative based on the singular value decomposition (SVD), where missing entries are handled naturally, and the number of underlying factors is imposed to be small and inferred in the light of observed entries. Although we assume missing at random, the proposed model is generalized to incorporate auxiliary information including missingness features. We also make a first attempt in the matrix-completion community to acquire new entries actively. By introducing a probit link function, we are able to handle counting matrices with the decomposed low-rank matrices latent. The basic model and its extensions are validated on</p><p>synthetic data, a movie-rating benchmark and a new data set presented for the first time.</p> / Dissertation

One view to cognition is that the symbol manipulating brain interprets the symbols of language based on the sensori-motor experiences of the agent. Such symbols, for example, what we refer to as nouns and verbs, are generalizations that the agent discovers through interactions with the environment. Given that an important subset of nouns correspond to objects (and object concepts), in this thesis, how function and appearance-based object concepts can be created through affordances has been studied. For this, a computational system, which is able to create object concepts through simple interactions with the objects in the environment, is proposed. Namely, the robot applies a set of built-in behaviors (such as pushing, lifting, grasping) on a set of objects to learn their aordances, through which objects affording similar functions are grouped into object concepts. Moreover, the thesis demonstrates that the discovered object concepts are beneficial for learning new tasks by analyzing the learning performance of learning a new task with and without object concepts.

Multi-task learning refers to learning multiple tasks simultaneously, in order to avoid tabula rasa learning and to share information between similar tasks during learning. We consider a multi-task Gaussian process regression model that learns related functions by inducing correlations between tasks directly. Using this model as a reference for three other multi-task models, we provide a broad unifying view of multi-task learning. This is possible because, unlike the other models, the multi-task Gaussian process model encodes task relatedness explicitly. Each multi-task learning model generally assumes that learning multiple tasks together is beneficial. We analyze how and the extent to which multi-task learning helps improve the generalization of supervised learning. Our analysis is conducted for the average-case on the multi-task Gaussian process model, and we concentrate mainly on the case of two tasks, called the primary task and the secondary task. The main parameters are the degree of relatedness ρ between the two tasks, and πS, the fraction of the total training observations from the secondary task. Among other results, we show that asymmetric multitask learning, where the secondary task is to help the learning of the primary task, can decrease a lower bound on the average generalization error by a factor of up to ρ2πS. When there are no observations for the primary task, there is also an intrinsic limit to which observations for the secondary task can help the primary task. For symmetric multi-task learning, where the two tasks are to help each other to learn, we find the learning to be characterized by the term πS(1 − πS)(1 − ρ2). As far as we are aware, our analysis contributes to an understanding of multi-task learning that is orthogonal to the existing PAC-based results on multi-task learning. For more than two tasks, we provide an understanding of the multi-task Gaussian process model through structures in the predictive means and variances given certain configurations of training observations. These results generalize existing ones in the geostatistics literature, and may have practical applications in that domain. We evaluate the multi-task Gaussian process model on the inverse dynamics problem for a robot manipulator. The inverse dynamics problem is to compute the torques needed at the joints to drive the manipulator along a given trajectory, and there are advantages to learning this function for adaptive control. A robot manipulator will often need to be controlled while holding different loads in its end effector, giving rise to a multi-context or multi-load learning problem, and we treat predicting the inverse dynamics for a context/load as a task. We view the learning of the inverse dynamics as a function approximation problem and place Gaussian process priors over the space of functions. We first show that this is effective for learning the inverse dynamics for a single context. Then, by placing independent Gaussian process priors over the latent functions of the inverse dynamics, we obtain a multi-task Gaussian process prior for handling multiple loads, where the inter-context similarity depends on the underlying inertial parameters of the manipulator. Experiments demonstrate that this multi-task formulation is effective in sharing information among the various loads, and generally improves performance over either learning only on single contexts or pooling the data over all contexts. In addition to the experimental results, one of the contributions of this study is showing that the multi-task Gaussian process model follows naturally from the physics of the inverse dynamics.

Transfer Learning is an emerging framework for learning from data that aims at intelligently transferring information between tasks. This is achieved by developing algorithms that can perform multiple tasks simultaneously, as well as translating previously acquired knowledge to novel learning problems. In this thesis, we investigate the application of Gaussian Processes to various forms of transfer learning with a focus on classification problems. This process initiates with a thorough introduction to the framework of Transfer learning, providing a clear taxonomy of the areas of research. Following that, we continue by reviewing the recent advances on Multi-task learning for regression with Gaussian processes, and compare the performance of some of these methods on a real data set. This review gives insights about the strengths and weaknesses of each method, which acts as a point of reference to apply these methods to other forms of transfer learning. The main contributions of this thesis are reported in the three following chapters. The third chapter investigates the application of Multi-task Gaussian processes to classification problems. We extend a previously proposed model to the classification scenario, providing three inference methods due to the non-Gaussian likelihood the classification paradigm imposes. The forth chapter extends the multi-task scenario to the semi-supervised case. Using labeled and unlabeled data, we construct a novel covariance function that is able to capture the geometry of the distribution of each task. This setup allows unlabeled data to be utilised to infer the level of correlation between the tasks. Moreover, we also discuss the potential use of this model to situations where no labeled data are available for certain tasks. The fifth chapter investigates a novel form of transfer learning called meta-generalising. The question at hand is if, after training on a sufficient number of tasks, it is possible to make predictions on a novel task. In this situation, the predictor is embedded in an environment of multiple tasks but has no information about the origins of the test task. This elevates the concept of generalising from the level of data to the level of tasks. We employ a model based on a hierarchy of Gaussian processes, in a mixtures of expert sense, to make predictions based on the relation between the distributions of the novel and the training tasks. Each chapter is accompanied with a thorough experimental part giving insights about the potentials and the limits of the proposed methods.

Apprendre des tâches simultanément peut améliorer la performance de prédiction par rapport à l'apprentissage de ces tâches de manière indépendante. Dans cette thèse, nous considérons l'apprentissage multi-tâche lorsque le nombre de tâches est grand. En outre, nous débattons des restrictions imposées sur les tâches. Ces restrictions peuvent être trouvées dans les méthodes de l'état de l'art. Plus précisément on trouve les restrictions suivantes : l'imposition du même espace d'étiquette sur les tâches, l'exigence des mêmes exemples d'apprentissage entre tâches et / ou supposant une hypothèse de corrélation globale entre tâches. Nous proposons des nouveaux classificateurs multi-tâches qui relaxent les restrictions précédentes. Nos classificateurs sont considérés en fonction de la théorie de l'apprentissage PAC des classifieurs faibles, donc, afin de parvenir à un faible taux d'erreur de classification, un ensemble de ces classifieurs faibles doivent être appris. Ce cadre est appelé l'apprentissage d'ensembles, dans lequel nous proposons un algorithme d'apprentissage multi-tâche inspiré de l'algorithme Adaboost pour seule tâche. Différentes variantes sont proposées également, à savoir, les forêts aléatoires pour le multi-tâche, c'est une méthode d'apprentissage d'ensemble, mais fondée sur le principe statistique d'échantillonnage Bootstrap. Enfin, nous donnons une validation expérimentale qui montre que l'approche sur-performe des méthodes existantes et permet d'apprendre des nouvelles configurations de tâches qui ne correspondent pas aux méthodes de l'état de l'art. / Learning multiple related tasks jointly by exploiting their underlying shared knowledge can improve the predictive performance on every task compared to learning them individually. In this thesis, we address the problem of multi-task learning (MTL) when the tasks are heterogenous: they do not share the same labels (eventually with different number of labels), they do not require shared examples. In addition, no prior assumption about the relatedness pattern between tasks is made. Our contribution to multi-task learning lies in the framework of en- semble learning where the learned function consists normally of an ensemble of "weak " hypothesis aggregated together by an ensemble learning algorithm (Boosting, Bagging, etc.). We propose two approaches to cope with heterogenous tasks without making prior assumptions about the relatedness patterns. For each approach, we devise novel multi-task weak hypothesis along with their learning algorithms then we adapt a boosting algorithm to the multi-task setting. In the first approach, the weak classi ers we consider are 2-level decision stumps for di erent tasks. A weak classi er assigns a class to each instance on two tasks and abstain on other tasks. The weak classi ers allow to handle dependencies between tasks on the instance space. We introduce di fferent effi cient weak learners. We then consider Adaboost with weak classi ers which can abstain and adapt it to multi-task learning. In an empirical study, we compare the weak learners and we study the influence of the number of boosting rounds. In the second approach, we develop the multi-task Adaboost environment with Multi-Task Decision Trees as weak classi ers. We fi rst adapt the well known decision tree learning to the multi-task setting. We revise the information gain rule for learning decision trees in the multi-task setting. We use this feature to develop a novel criterion for learning Multi-Task Decision Trees. The criterion guides the tree construction by learning the decision rules from data of di fferent tasks, and representing diff erent degrees of task relatedness. We then modify MT-Adaboost to combine Multi-task Decision Trees as weak learners. We experimentally validate the advantage of our approaches; we report results of experiments conducted on several multi-task datasets, including the Enron email set and Spam Filtering collection.

京都大学 / 新制・課程博士 / 博士(情報学) / 甲第23540号 / 情博第770号 / 新制||情||131(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)准教授 吉井 和佳, 教授 河原 達也, 教授 西野 恒, 教授 鹿島 久嗣 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM