Gibbs-randomness proposition for energy efficient deep learning