Mengyuan Zhao 2015-09-07
来自cslt Wiki
- reproduce batch-normalization experiment on aurora4 dataset. get similar result with ZZY
- test Ad-grd and Ad-Max on sino 100h cn dataset. Ad-Max's result is not good. Ad-grd looks reasonable(tr_acc, cv_acc is ok), but training diverge after 5 iters.
- fix the bug of tag-lm while doing biglm decoding with multi tag.