Update of 4/1/2022: Just finished reading this paper. It sounds very simple. You wonder why nobody else came up with their solution before. Let's wait and see how the machine learning community receives this paper. It is apparently under review. The senior author of this paper, i.e. Philipp H. S. Torr is a highly cited researcher (professor at the University of Oxford). Some of his collaborators in the past are also well known like Andrea Vedaldi and Andrew Zisserman.
This could potentially be a game changer for machine learning! It may cut the model training time in half! I have not yet studied this paper.
From the abstract:
"Using backpropagation to compute gradients of objective functions for optimization has remained a mainstay of machine learning. Backpropagation, or reverse-mode differentiation, is a special case within the general family of automatic differentiation algorithms that also includes the forward mode. We present a method to compute gradients based solely on the directional derivative that one can compute exactly and efficiently via the forward mode. We call this formulation the forward gradient, an unbiased estimate of the gradient that can be evaluated in a single forward run of the function, entirely eliminating the need for backpropagation in gradient descent. We demonstrate forward gradient descent in a range of problems, showing substantial savings in computation and enabling training up to twice as fast in some cases."
No comments:
Post a Comment