Momentumoptimointi

Momentumoptimointi, often translated as momentum optimization, is a technique used in machine learning and deep learning to speed up the convergence of training algorithms. It builds upon standard stochastic gradient descent (SGD) by incorporating information from past gradients to influence the direction of the current update. Instead of solely relying on the immediate gradient of the loss function, momentum methods accumulate a "velocity" term that represents a moving average of past gradients.

The core idea behind momentum is to dampen oscillations that can occur in SGD, especially in directions where the gradient frequently changes sign or magnitude. By averaging past updates, momentum helps the optimization process to maintain progress in consistent directions and smooth out erratic movements. This can lead to faster convergence to a local or global minimum of the loss function.

A common implementation of momentum involves a decay factor, often denoted by beta (β), which controls how much of the previous velocity is retained in the current update. A higher beta value means more emphasis on past gradients. The update rule typically involves calculating the current gradient, updating the velocity by combining the current gradient with the previous velocity (scaled by beta), and then using this updated velocity to adjust the model's parameters. This approach can be particularly beneficial for training deep neural networks where the loss landscape can be complex and prone to getting stuck in shallow local minima or plateaus.