Graph-based latent space analysis
Title: Graph-based latent space analysis
Abstract: In recent years, Deep Neural Networks (DNNs) have achieved state of the art performance in a vast range of machine learning tasks via end-to-end learning. Unfortunately, there are various undesirable consequences from end-to-end learning, such as i) difficulty to ensure its generalization capabilities, ii) robustness to small perturbations of inputs, iii) ensuring latent space consistency. In this talk, we will tackle these three problems via the analysis of the latent spaces of DNNs. While analyzing these spaces directly is difficult due to the high dimensionality of representations and the stochasticity of the training process, we propose the use of graphs to represent the geometry of these spaces. In more details, we will first show how measures derived from Graph Signal Processing (GSP) are directly linked to the generalization (i) and robustness (ii) of the DNNs, and then use the results from this analysis to improve the performance of DNNs in supervised classification and visual localization (iii).
Web page: https://cadurosar.github.io
Talk online : on https://lpensl.my.webex.com/
Room : Webconf3