Date(s) : 06/03/2020 iCal
14 h 00 min - 15 h 00 min
We define maximum-entropy models of stationary processes from non-linear representations. To capture coherent structures in a random process such as Turbulent flows, we focus on the problemof choosing sufficient statistical moments. It amounts to find features corresponding to the statistical patterns of coherent structures. A new set of covariance moments are learnt-by-hand over the last few years. It brings about a conceptual connection betweenthe rectifier non-linearity in convolutional neural networks and the phase in Fourier and wavelet representations. We conjecture that structures in the phase play a key role for pattern recognition and high-dimensional data modeling.