Information contraction in noisy binary neural networks and its implications