Nonlinear ICA using Time-contrastive learning

July 31, 2024

• We learn features that enable discriminating data points from different time segments
• We use ordinary neural network training: Last hidden layer gives the features
• Surprising theoretical result: Estimates a nonlinear ICA model
  • with general nonlinear mixing $x(t) = f(s(t))$
  • nonstationary components $s_{i}(t)$