Forêt d'arbres décisionnels

Résumé

vignette|Illustration du principe de construction d'une forêt aléatoire comme agrégation d'arbre aléatoires. En apprentissage automatique, les forêts d'arbres décisionnels (ou forêts aléatoires de l'anglais random forest classifier) forment une méthode d'apprentissage ensembliste. Ils ont été premièrement proposées par Ho en 1995 et ont été formellement proposées en 2001 par Leo Breiman et Adele Cutler. Cet algorithme combine les concepts de sous-espaces aléatoires et de bagging. L'algorithme des forêts d'arbres décisionnels effectue un apprentissage sur de multiples arbres de décision entraînés sur des sous-ensembles de données légèrement différents. Algorithme La base du calcul repose sur l'apprentissage par arbre de décision. La proposition de Breiman vise à corriger plusieurs inconvénients connus de la méthode initiale, comme la sensibilité des arbres uniques à l'ordre des prédicteurs, en calculant un ensemble de {B} arbres partielleme

Source officielle

https://fr.wikipedia.org/wiki/For%C3%AAt_d'arbres_d%C3%A9cisionnels

À propos de ce résultat

Cette page est générée automatiquement et peut contenir des informations qui ne sont pas correctes, complètes, à jour ou pertinentes par rapport à votre recherche. Il en va de même pour toutes les autres pages de ce site. Veillez à vérifier les informations auprès des sources officielles de l'EPFL.

Publications associées (57)

Publications associées

Chargement

Personnes associées (5)

Personnes associées

Chargement

Unités associées (6)

Unités associées

Chargement

Concepts associés (15)

Concepts associés

Chargement

Cours associés (57)

Students understand basic concepts and methods of machine learning. They can describe them in mathematical terms and can apply them to data using a high-level programming language (julia/python/R).

This course teaches the basic techniques, methodologies, and practical skills required to draw meaningful insights from a variety of data, with the help of the most acclaimed software tools in the data science world (pandas, scikit-learn, Spark, etc.)

A course on statistical machine learning for supervised and unsupervised learning

Afficher plus

Cours associés

Chargement

Séances de cours associées (159)

Séances de cours associées

Chargement

Theory of Deep Learning: Neural Tangent Kernel and Beyond

Arthur Ulysse Jacot-Guillarmod

In the recent years, Deep Neural Networks (DNNs) have managed to succeed at tasks that previously appeared impossible, such as human-level object recognition, text synthesis, translation, playing games and many more. In spite of these major achievements, our understanding of these models, in particular of what happens during their training, remains very limited. This PhD started with the introduction of the Neural Tangent Kernel (NTK) to describe the evolution of the function represented by the network during training. In the infinite-width limit, i.e. when the number of neurons in the layers of the network grows to infinity, the NTK converges to a deterministic and time-independent limit, leading to a simple yet complete description of the dynamics of infinitely-wide DNNs. This allowed one to give the first general proof of convergence of DNNs to a global minimum, and yielded the first description of the limiting spectrum of the Hessian of the loss surface of DNNs throughout training.More importantly, the NTK plays a crucial role in describing the generalization abilities of DNNs, i.e. the performance of the trained network on unseen data. The NTK analysis uncovered a direct link between the function learned by infinitely wide DNNs and Kernel Ridge Regression predictors, whose generalization properties are studied in this thesis using tools of random matrix theory. Our analysis of KRR reveals the importance of the eigendecomposition of the NTK, which is affected by a number of architectural choices. In very deep networks, an ordered regime and a chaotic regime appear, determined by the choice of non-linearity and the balance between the weights and bias parameters; these two phases are characterized by different speeds of decay of the eigenvalues of the NTK, leading to a tradeoff between convergence speed and generalization. In practical contexts such as Generative Adversarial Networks or Topology Optimization, the network architecture can be chosen to guarantee certain properties of the NTK and its spectrum.These results give an almost complete description DNNs in this infinite-width limit. It is then natural to wonder how it extends to finite-width networks used in practice. In the so-called NTK regime, the discrepancy between finite- and infinite-widths DNNs is mainly a result of the variance w.r.t. to the sampling of the parameters, as shown empirically and mathematically relying on the similarity between DNNs and random feature models.In contrast to the NTK regime, where the NTK remains constant during training, there exist so-called active regimes, where the evolution of the NTK is significant, which appear in a number of settings. One such regime appears in Deep Linear Networks with a very small initialization, where the training dynamics approach a sequence of saddle-points, representing linear maps of increasing rank, leading to a low-rank bias which is absent in the NTK regime.

EPFL2022

Chargement

Prediction of patient-reported physical activity scores from wearable accelerometer data: a feasibility study

Inès Bahej, Martin Jaggi

Many diseases are characterized by limitations in mobility, including a wide range of musculoskeletal and neurological conditions. Reduced mobility impacts a patients ability to perform activities of daily living, which in turn reduces health-related quality of life. Mobility can be assessed by collecting patient-reported outcome scores from standardized questionnaires and by directly measuring physical activity parameters from wearable accelerometer data. In this work, we explored the relationship between subjectively and objectively measured mobility by training machine learning models to predict patient responses based on features derived from real-world acceleration data. Our method achieved up to 82% accuracy using a random forest classifier and set the basis to develop novel data-driven digital biomarkers for objective, quantitative and more frequent evaluation of patients’ mobility.

2018

Chargement

In this thesis, I focus on monitoring of patients suffering from cardiovascular and neurological diseases through the use of wearable devices. The main diseases considered in this thesis are atrial fibrillation (AF), myocardial infarction (MI), and epilepsy. The proposed methods for the detection of the considered cardiovascular diseases use ECG as the main biosignal, whereas the main biosignal used for detecting epileptic seizures is EEG. Firstly, I propose a hierarchical heart-rhythm classification method for AF detection from a single lead ECG recording. The proposed method is based on the features that capture the morphology of important ECG signal segments, along with heart-rate oscillations and time and frequency domain features of the ECG signal. Furthermore, the classification scheme used in this method incorporates two different classifiers: a multiclass classifier and a random forest classifier, resulting in an F1 score of 78.95% for AF detection. Secondly, I address the problem of early detection and prevention of MI by introducing a real-time event-driven classification technique that uses a hierarchical classifier with multiple levels. The main goal of this technique is to maintain a high classification accuracy while reducing the complexity of the classification algorithm. The reduction in computational complexity, in turn, results in a longer battery life, which is an important factor for wearable devices. The proposed technique is validated on a public database and ported on a real-life wearable device. The experimental evaluation of the proposed technique on MI data shows that this scheme reduces the energy consumption by a factor of 2.60, with no significant loss in classification performance. The third chapter of this thesis is split into two parts and it focuses on epilepsy. In the first part, I present a real-time method for personalized epileptic seizure detection that uses EEG signals acquired from two electrode pairs: F7T3, and F8T4. The proposed method reaches a sensitivity of 90.98% and specificity of 92.10% on the database used in this study. Furthermore, this method is ported on a pair of eyeglasses in which the used electrodes are embedded and hidden in the temples, allowing for 2.71 days of operation on a single battery charge. This method overcomes the lack of portability and the effect of social stigma of EEG caps, which are used as a gold standard technique for epilepsy detection. Since the main pitfall of epilepsy detection algorithms is the unacceptably high number of false alarms, in the last part of my thesis, I propose an interpretable patient-specific approach to false alarm reduction for epilepsy detection. This approach is based on similarly occurring morphological EEG signal patterns (seizure signature) that occur frequently during seizures. The proposed approach has been experimentally validated on more than 5500 hours of long-term EEG recordings, resulting in a high classification performance with no false positive alarms. The high degree of interpretability of this method can help physicians discover seizures faster, as well as it can be used to improve data labeling quality in publicly available databases, which confirms its applicability for long-term seizure monitoring.

EPFL2020

Chargement

Theory of Deep Learning: Neural Tangent Kernel and Beyond

Arthur Ulysse Jacot-Guillarmod

EPFL2022

EPFL2020