Do you also have this one very reasonable friend who always slows you down when you have a crazy idea, like opening a bar, for example? Dampening is this friend to momentum.

Dampening makes sure that the optimizer isn't taking too big steps on the loss landscape what might happen if you use momentum only. Finally, the higher the gradient is, the more dampening reduces the step size.

Dampening is not well researched yet, so it's not really proven and mostly set to 0 as default. But it may be interesting to play around with.

Boost model performance quickly with AI-powered labeling and 100% QA.

Learn more
Last modified