Automatic Learning Rate Maximization in Large Adaptive Machines