Prolégomènes aux réseaux de neurones multirésolution
Vincent Lostanlen · HAL (Le Centre pour la Communication Scientifique Directe) · 2020
Deep convolutional networks (convnets) have a predominant importance in pattern recognition, and in machine listening in particular. Although the resort to two-dimensional convolutions in the time-frequency domain has led to successful results in audio classification, the prospect of employing them directly in the raw waveform, without any a priori on the underlying time-frequency representation,would allow to take full advantage of the "end-to-end learning" paradigm. However, applying statistical machine learning to yield a spectrogram-like time-frequency representation raises issues of aliasing, scalability, and phase reconstruction. In this document, I address those issues by initiating a research program at the intersection of multiresolution approximation (MRA) and deep learning. I outline the main theoretical and experimental challenges that remain to be solved in order to train multiresolution neural networks (MuReNN) on machine listening tasks, either supervised or unsupervised.