Extreme learning machine | EPFL Graph Search

En apprentissage automatique, le terme extreme learning machine (machine à apprentissage extrême) fait référence à un type de réseau de neurones. Sa spécificité est de n'avoir qu'une seule couche de nœuds cachés, où les poids des entrées de connexion de nœuds cachés sont répartis au hasard et jamais mis à jour. Ces poids entre les nœuds cachés d'entrée et les sorties sont appris en une seule étape, ce qui revient essentiellement à l'apprentissage d'un modèle linéaire. Le nom "extreme learning machine" (ELM) a été donné à ces modèles par Guang-Bin Huang, mais le principe était déjà connu. Ces modèles peuvent produire une bonne performance de généralisation et avoir un processus d'apprentissage beaucoup plus rapide que les réseaux entraînés en utilisant la rétropropagation du gradient. Algorithme Le plus simple algorithme d'apprentissage ELM a un modèle de la forme : \mathbf{\hat{Y}} = \mathbf{W}_2 \sigma(\mathbf{W}_1 x) où est la matrice des pondérations d'en

MonStereo: When Monocular and Stereo Meet at the Tail of 3D Human Localization

Alexandre Massoud Alahi, Lorenzo Bertoni, Sven Kreiss, Taylor Ferdinand Mordan

Monocular and stereo visions are cost-effective solutions for 3D human localization in the context of self-driving cars or social robots. However, they are usually developed independently and have their respective strengths and limitations. We propose a novel unified learning framework that leverages the strengths of both monocular and stereo cues for 3D human localization. Our method jointly (i) associates humans in left-right images, (ii) deals with occluded and distant cases in stereo settings by relying on the robustness of monocular cues, and (iii) tackles the intrinsic ambiguity of monocular perspective projection by exploiting prior knowledge of the human height distribution. We specifically evaluate outliers as well as challenging instances, such as occluded and far-away pedestrians, by analyzing the entire error distribution and by estimating calibrated confidence intervals. Finally, we critically review the official KITTI 3D metrics and propose a practical 3D localization metric tailored for humans.

2021

Deep learning-based analysis of multiple sclerosis lesions with high and ultra-high field MRI

Francesco La Rosa

Multiple sclerosis (MS) is the most common demyelinating disease of the central nervous system and affects almost 3 million people worldwide. There is currently no cure for MS, and its symptoms, starting with fatigue and weakness, often progress over time into disabilities and cognitive impairments. Hence, early diagnosis of the disease is crucial to start appropriate treatment and delay MS progression. Diagnosis, prognosis, and follow-up of MS patients rely mostly on magnetic resonance imaging (MRI) findings such as white matter lesions. This conventional biomarker, however, lacks specificity in differentiating MS from mimic conditions and does not strongly correlate with disability progression. Recently, advanced imaging biomarkers, including cortical lesions and paramagnetic rim lesions, have emerged as being specific to MS and predicting disease severity. Nevertheless, their assessment requires highly specialized experts and is extremely time-consuming. Thus, there is currently an unmet need for decision support tools to automate the analysis of these biomarkers that can ultimately enhance the understanding of the disease and improve patient care.The present thesis describes the development and validation of different deep learning approaches for the automated analysis of MS lesions at high and ultra-high field MRI, focusing on cortical and paramagnetic rim lesions. The main contributions are the following. First, an automated method for the segmentation of cortical and white matter lesions with only two 3T MRI sequences is proposed. Based on a patch-wise convolutional neural network, this method is validated on a large multi-site cohort of MS patients, outperforming existing machine learning approaches. Second, we translated our approach to ultra-high field MRI, exploring specialized MRI contrasts. Cortical Lesion AI-based assessment in MS (CLAIMS) is a fully-automated framework providing cortical lesion detection and classification at 7T MRI. Third, the issue of missing MRI modalities was tackled. As the MP2RAGE MRI sequence is highly desirable to detect cortical lesions, yet not often acquired in clinical practice, a generative adversarial network is introduced to synthesize MP2RAGE images from MPRAGE. This approach generates realistic MP2RAGE images which, compared to the MPRAGE, significantly improve the lesion and brain tissue segmentation masks computed by automated tools. Fourth, the first deep learning-based approach for the classification of paramagnetic rim lesions is proposed (RimNet). RimNet has been evaluated with different MRI contrasts at 3T and 7T, and its performance is close to that of an expert.To summarize, this thesis aims at providing robust deep learning-based approaches for the automated analysis of MS lesions. All methods developed have been designed to meet specific clinical needs. A minimal number of clinically relevant MRI sequences was proven sufficient to outperform existing techniques and approach the experts' manual evaluation. As these methods generalize well to multi-site datasets at both high and ultra-high field MRI, they represent the first step toward an automated and standardized assessment of advanced MS biomarkers.

EPFL2022

Deep learning-based analysis of multiple sclerosis lesions with high and ultra-high field MRI

Francesco La Rosa

EPFL2022