Classification of fall directions via wearable motion sensors

Effective fall-detection and classification systems are vital in mitigating severe medical and economical consequences of falls to people in the fall risk groups. One class of such systems is based on wearable sensors. While there is a vast amount of academic work on this class of systems, not much effort has been devoted to the investigation of effective and robust algorithms and like-for-like comparison of state-ofthe-art algorithms using a sufficiently large dataset. In this article, fall-direction classification algorithms are presented and compared on an extensive dataset, comprising a total of 1600 fall trials. Eight machine learning classifiers are implemented for fall-direction classification into four basic directions (forward, backward, right, and left). These are, namely, Bayesian decision making (BDM), least squares method (LSM), k-nearest neighbor classifier (k-NN), artificial neural networks (ANNs), support vector machines (SVMs), decision-tree classifier (DTC), random forest (RF), and adaptive boosting or AdaBoost (AB). BDM achieves perfect classification, followed by k-NN, SVM, and RF. Data acquired from only a single motion sensor unit, worn at the waist of the subject, are processed for experimental verification. Four of the classifiers (BDM, LSM, k-NN, and ANN) are modified to handle the presence of data from an unknown class and evaluated on the same dataset. In this robustness analysis, ANN and k-NN yield accuracies above 96.2%. The results obtained in this study are promising in developing real-world fall-classification systems as they enable fast and reliable classification of fall directions.(c) 2021 Elsevier Inc. All rights reserved.

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

Adaptive wake and sleep detection for wearable systems

Walter Karlen

Sleep problems and disorders have a serious impact on human health and wellbeing. The rising costs for treating sleep-related chronic diseases in industrialized countries demands efficient prevention. Low-cost, wearable sleep / wake detection systems which give feedback on the wearer's "sleep performance" are a promising approach to reduce the risk of developing serious sleep disorders and fatigue. Not all bio-medical signals that are useful for sleep / wake discrimination can be easily recorded with wearable systems. Sensors often need to be placed in an obtrusive location on the body or cannot be efficiently embedded into a wearable frame. Furthermore, wearable systems have limited computational and energetic resources, which restrict the choice of sensors and algorithms for online processing and classification. Since wearable systems are used outside the laboratory, the recorded signals tend to be corrupted with additional noise that influences the precision of classification algorithms. In this thesis we present the research on a wearable sleep / wake classifier system that relies on cardiorespiratory (ECG and respiratory effort) and activity recordings and that works autonomously with minimal user interaction. This research included the selection of optimal signals and sensors, the development of a custom-tailored hardware demonstrator with embedded classification algorithms, and the realization of experiments in real-world environments for the customization and validation of the system. The processing and classification of the signals were based on Fourier transformations and artificial neural networks that are efficiently implementable into digital signal controllers. Literature analysis and empiric measurements revealed that cardiorespiratory signals are more promising for a wearable sleep / wake classification than clinically used signals such as brain potentials. The experiments conducted during this thesis showed that inter-subject differences within the recorded physiological signals make it difficult to design a sleep / wake classification model that can generalize to a group of subjects. This problem was addressed in two ways: First by adding features from another signal to the classifier, that is, measuring the behavioral quiescence during sleep using accelerometers. Conducted research on different feature extraction methods from accelerometer data showed that this data generalizes well for distinct subjects in the study group. In addition, research on user-adaptation methods was conducted. Behavioral sleep and wake measures, notably the measurement of reactivity and activity, were developed to build up a priori knowledge that was used to adapt the classification algorithm automatically to new situations. This thesis demonstrates the design and development of a low-cost, wearable hardware and embedded software for on-line sleep / wake discrimination. The proposed automatic user-adaptive classifier is advantageous compared to previously suggested classification methods that generalize over multiple subjects, because it can take changes in the wearer's physiology and sleep / wake behavior into account without adjustment from a human expert. The results of this thesis contribute to the development of smart, wearable, bio-physiological monitoring systems which require a high degree of autonomy and have only low computational resources available. We believe that the proposed sleep / wake classification system is a first promising step toward a context-aware system for sleep management, sleep disorder prevention, and reduction of fatigue.

EPFL2009

Predicting in Uncertain Environments: Methods for Robust Machine Learning

Paul Thierry Yves Rolland

One of the main goal of Artificial Intelligence is to develop models capable of providing valuable predictions in real-world environments. In particular, Machine Learning (ML) seeks to design such models by learning from examples coming from this same environment. However, the real world is most of the time not static, and the environment in which the model will be used can differ from the one in which it is trained. It is hence desirable to design models that are robust to changes of environments. This encapsulates a large family of topics in ML, such as adversarial robustness, meta-learning, domain adaptation and others, depending on the way the environment is perturbed.In this dissertation, we focus on methods for training models whose performance does not drastically degrade when applied to environments differing from the one the model has been trained in. Various types of environmental changes will be treated, differing in their structure or magnitude. Each setup defines a certain kind of robustness to certain environmental changes, and leads to a certain optimization problem to be solved. We consider 3 different setups, and propose algorithms for solving each associated problem using 3 different types of methods, namely, min-max optimization (Chapter 2), regularization (Chapter 3) and variable selection (Chapter 4).Leveraging the framework of distributionally robust optimization, which phrases the problem of robust training as a min-max optimization problem, we first aim to train robust models by directly solving the associated min-max problem. This is done by exploiting recent work on game theory as well as first-order sampling algorithms based on Langevin dynamics. Using this approach, we propose a method for training robust agents in the scope of Reinforcement Learning.We then treat the case of adversarial robustness, i.e., robustness to small arbitrary perturbation of the model's input. It is known that neural networks trained using classical optimization methods are particularly sensitive to this type of perturbations. The adversarial robustness of a model is tightly connected to its smoothness, which is quantified by its so-called Lipschitz constant. This constant measures how much the model's output changes upon any bounded input perturbation. We hence develop a method to estimate an upper bound on the Lipschitz constant of neural networks via polynomial optimization, which can serve as a robustness certificate against adversarial attacks. We then propose to penalize the Lipschitz constant during training by minimizing the 1-path-norm of the neural network, and we develop an algorithm for solving the resulting regularized problem by efficiently computing the proximal operator of the 1-path-norm term, which is non-smooth and non-convex.Finally, we consider a scenario where the environmental changes can be arbitrary large (as opposed to adversarial robustness), but need to preserve a certain causal structure. Recent works have demonstrated interesting connections between robustness and the use of causal variables. Assuming that certain mechanisms remain invariant under some change of the environment, it has been shown that knowing the underlying causal structure of the data at hand allows to train models that are invariant to such changes. Unfortunately, in many cases, the causal structure is unknown. We thus propose a causal discovery algorithm from observational data in the case of non-linear additive model.

EPFL2022