machine learning - Explanation for Coordinate Descent and Subgradient -
how easy explanation of coordinate descent , subgradient solution in context of lasso.
an intuitive explanation followed proof helpful.
suppose have multivariate function f(w) k number of variables/parameters w (w_1, w_2, w_3, ..., w_k). parameters knobs , goal change these knobs in way f minimized function f. coordinate descent greedy method sense on each iteration change values of parameters w_i minimize f. easy implement , gradient descent guaranteed minimize f on each iteration , reach local minima.
picture borrowed internet through bing image search
as shown in picture above, function f has 2 parameters x , y. on each iteration either both of parameters changed fixed value c , value of function evaluated @ new point. if value higher , goal minimize function, change reversed selected parameter. same procedure done second parameter. 1 iteration of algorithm.
an advantage of using coordinate descent in problems computing gradient of function expensive.
sources

Comments
Post a Comment