Health Tech Lunch - Tomasini Umberto - “How deep convolutional neural networks lose spatial information with training - and how this influences explainable medical AI”
![Discover the Institute of Computer Science at HES-SO Valais-Wallis, a leader in software engineering and digital transformation. Our team of over 100 specialists and researchers enhances business competitiveness and quality of life through applied research projects. Explore our research areas in eHealth, energy and digitalization, and digital industry. Join our community committed to technological innovation.](/media/image/20/xlarge_2_1/header-iig-e-1378.jpg?3a3f65fa325395fb10883eef3935d9bb)
The Axe Santé is organising a Health Tech Lunch on Tuesday, May 23th from 12:00 to 12:45 at Swiss Digital Center in Sierre (Room Foyer Conf. Les Quilles»)
Biography : Umberto Tomasini started his PhD at EPFL in Physics and Machine Learning in 2020. He is currently interested in understanding how deep learning can learn and represent structured data, by theoretical and empirical means. Previously, he studied Statistical Physics at the University of Padova, and he was part of its Excellence School. In 2022 he won a Dean’s Award for excellence in teaching.
Abstract : A central question of machine learning is how deep nets manage to learn tasks in high dimensions. An appealing hypothesis is that they achieve this feat by building a representation of the data where information irrelevant to the task is lost. For image datasets, this view is supported by the observation that after (and not before) training, the neural representation becomes less and less sensitive to local distortions (diffeomorphisms) acting on images as the signal propagates through the net. This loss of sensitivity correlates with performance and surprisingly correlates with a gain of sensitivity to white noise acquired during training. We argue that two mechanisms, spatial and channel pooling, are being learnt by Convolutional Neural Networks to build a representation nearly invariant to smooth transformations. We develop an empirical procedure to disentangle these two effects. Finally, we introduce a toy model of data that captures our salient observations, and that can be treated analytically. In particular, this analysis explains why a succession of spatial pooling and non-linear operations can increase the sensitivity to random noise added to images.
Programme:
Registration : https://doodle.com/meeting/participate/id/e1GLzr0b