In this paper, we propose a simple, fast and easy to implement algorithm lossgrad (locally optimal step-size in gradient descent), which au- tomatically modifies the step-size in gradient descent during neural networks training. Given a function f, a point x, and the gradient rxf of f, we aim to nd the step-size h which is (locally) optimal, i.e. satisfies: h = arg min t0 f(x trxf): Making use of quadratic approximation, we show that the algorithm satisfies the above assumption. We experimentally show that our method is insensitive to the choice of initial learning rate while achieving results comparable to other methods.
JavaScript jest wyłączony w Twojej przeglądarce internetowej. Włącz go, a następnie odśwież stronę, aby móc w pełni z niej korzystać.