Learning in neural networks based on a generalized fluctuation theorem.