Analysis of neural networks. Applications to interpretability and uncertainty

Abstract

[en] From image creation and pattern recognition to speech and text processing, the outstanding performance of neural networks in a wide variety of fields has made them a popular tool among researchers. However, the fact that we do not fully understand why their performance is so successful or how they operate converts this technology into a black-box model based on trial and error. In this work, we attempt to give deep neural networks a mathematical representation and present different examples and applications that bring light to the understanding of neural networks’ behaviour and usage.