Target-Independent Active Learning via Distribution-Splitting