Quantifying many-body learning far from equilibrium with representation learning