Learning Stable Task Sequences from Demonstration with Linear Parameter Varying Systems and Hidden Markov Models

The problem of acquiring multiple tasks from demonstration is typi- cally divided in two sequential processes: (1) the segmentation or identification of different subgoals/subtasks and (2) a separate learning process that parameterizes a control policy for each subtask. As a result, segmentation criteria typically neglect the characteristics of control policies and rely instead on simplified models. This paper aims for a single model capable of learning sequences of complex time-independent control policies that provide robust and stable behavior. To this end, we first present a novel and efficient approach to learn goal-oriented time independent motion models by estimating both attractor and dynamic behavior from data guaranteeing stability using linear parameter varying (LPV) systems. This method enables learning complex task sequences with hidden Markov models (HMMs), where each state/subtask is given by a stable LPV system and where transitions are most likely around the corresponding attractor. We study the dynamics of the HMM-LPV model and propose a motion generation method that guarantees the stability of task sequences. We validate our approach in two sets of demonstrated human motions

Scene image classification and segmentation with quantized local descriptors and latent aspect modeling

Pedro Manuel Da Silva Quelhas

The ever increasing number of digital images in both public and private collections urges on the need for generic image content analysis systems. These systems need to be capable to capture the content of images from both scenes and objects, in a compact way that allows for fast search and comparison. Modeling images based on local invariant features computed at interest point locations has proven in recent years to achieve such capabilities and to provide a robust and versatile way to perform wide-baseline matching and search for both scene and object images. In this thesis we explore the use of local descriptors for image representation in the tasks of scene and object classification, ranking, and segmentation. More specifically, we investigate the combined use of text modeling methods and local invariant features. Firstly, our work attempts to elucidate whether a text like bag-of-visterms representation (histogram of quantized local visual features) is suitable for scene and object classification, and whether some analogies between discrete scene representations and text documents exist. We further explore the bag-of-visterms approach in a fusion framework, combining texture and color information for natural scene classification. Secondly, we investigate whether unsupervised, latent space models can be used as feature extractors for the classification task and to discover patterns of visual co-occurrence. In this direction, we show that Probabilistic Latent Semantic Analysis (PLSA) generates a compact scene representation, discriminative for accurate classification, and more robust than the bagof-visterms representation when less labeled training data is available. Furthermore, we show through aspect-based image ranking experiments, the ability of PLSA to automatically extract visually meaningful scene patterns, making such representation useful for browsing image collections. Finally, we further explore the use of the latent aspect modeling in an image segmentation task. By extending the representation resulting from the latent aspect modeling, we are able to introduce contextual information for image segmentation that goes beyond the traditional regional contextual modeling found for instance in Markov Random Field approaches.

École Polytechnique Fédérale de Lausanne2007

Multilingual Training and Adaptation in Speech Recognition

Sibo Tong

State-of-the-art acoustic models for Automatic Speech Recognition (ASR) are based on Hidden Markov Models (HMM) and Deep Neural Networks (DNN) and often require thousands of hours of transcribed speech data during training. Therefore, building multilingual ASR systems or systems on a language with few resources is a challenging task. Multilingual training and cross-lingual adaptation are potential solutions. However, context-dependent states modeling creates difficulties for multilingual and cross-lingual ASR because of the large increase in context dependent labels arising from the phone set mismatch. The goal of this thesis is to improve current state-of-the-art acoustic modeling techniques in general for ASR, with a particular focus on multilingual ASR and cross-lingual adaptation. We systematically exploited new training frameworks, from Maximum Likelihood Estimation, Connectionist Temporal Classification to Maximum Mutual Information, in the context of phoneme-based multilingual training. In order to minimize the negative effects of data impurity arising from language mismatch, we investigated language adaptive training approaches which help further improve the multilingual ASR performance. Through comprehensive experimental comparison we demonstrated that phoneme-based multilingual models are easily extensible to unseen phonemes of new languages, from which the cross-lingual adaptation yields significant improvement over traditional approaches on limited data. Finally, we proposed a semi-supervised training approach based on dropout to boost the performance in low-resourced languages using untranscribed data. In the other part of the thesis, we conducted more theoretical analysis of techniques found to be useful in sequential multilingual training. More specifically, we revisited the recurrent architecture based on Bayesâs theorem. This leads to a Bayesian recurrent unit dictated by the probabilistic formulation and naturally support a backward recursion. Experiments show that the proposed architecture exceeds the performance of conventional recurrent network. Together, this thesis constitutes a thorough analysis of the current field. Through theoretical and experimental comparisons, the proposed approaches are shown to yield significant improvement over the conventional hybrid systems on multilingual speech recognition.

EPFL2020

Robust Adaptive Decision Making: Bayesian Optimization and Beyond

Ilija Bogunovic

The central task in many interactive machine learning systems can be formalized as the sequential optimization of a black-box function. Bayesian optimization (BO) is a powerful model-based framework for \emph{adaptive} experimentation, where the primary goal is the optimization of the black-box function via sequentially chosen decisions. In many real-world tasks, it is essential for the decisions to be \emph{robust} against, e.g., adversarial failures and perturbations, dynamic and time-varying phenomena, a mismatch between simulations and reality, etc. Under such requirements, the standard methods and BO algorithms become inadequate. In this dissertation, we consider four research directions with the goal of enhancing robust and adaptive decision making in BO and associated problems. First, we study the related problem of level-set estimation (LSE) with Gaussian Processes (GPs). While in BO the goal is to find a maximizer of the unknown function, in LSE one seeks to find all "sufficiently good" solutions. We propose an efficient confidence-bound based algorithm that treats BO and LSE in a unified fashion. It is effective in settings that are non-trivial to incorporate into existing algorithms, including cases with pointwise costs, heteroscedastic noise, and multi-fidelity setting. Our main result is a general regret guarantee that covers these aspects. Next, we consider GP optimization with robustness requirement: An adversary may perturb the returned design, and so we seek to find a robust maximizer in the case this occurs. This requirement is motivated by, e.g., settings where the functions during optimization and implementation stages are different. We propose a novel robust confidence-bound based algorithm. The rigorous regret guarantees for this algorithm are established and complemented with an algorithm-independent lower bound. We experimentally demonstrate that our robust approach consistently succeeds in finding a robust maximizer while standard BO methods fail. We then investigate the problem of GP optimization in which the reward function varies with time. The setting is motivated by many practical applications in which the function to be optimized is not static. We model the unknown reward function via a GP whose evolution obeys a simple Markov model. Two confidence-bound based algorithms with the ability to "forget" about old data are proposed. We obtain regret bounds for these algorithms that jointly depend on the time horizon and the rate at which the function varies. Finally, we consider the maximization of a set function subject to a cardinality constraint

k

in the case a number of items

\tau

from the returned set may be removed. One notable application is in batch BO where we need to select experiments to run, but some of them can fail. Our focus is on the worst-case adversarial setting, and we consider both \emph{submodular} (i.e., satisfies a natural notion of diminishing returns) and \emph{non-submodular} objectives. We propose robust algorithms that achieve constant-factor approximation guarantees. In the submodular case, the result on the maximum number of allowed removals is improved to

\tau = o(k)

in comparison to the previously known

\tau=o(\sqrt{k})

. In the non-submodular case, we obtain new guarantees in the support selection and batch BO tasks. We empirically demonstrate the robust performance of our algorithms in these, as well as, in data summarization and influence maximization tasks.

EPFL2019