PP: Soft-DTW: a differentiable loss function for time-series

Problem: new loss

Label: new loss; 

Abstract:

A differentiable learning loss; 

Introduction:

supervised learning: learn a mapping that links an input to an output object. 

output object is a time series.  

Prediction: two multi-layer perceptrons, the first use Euclidean loss and the second use soft-DTW as a loss function. --------> soft-DTW, better sharp changes. 

DTW computes the best possible alignment between two time series.