Setting starting values for trainable parameters using information from model gradients to improve convergence and final performance.