diff options
Diffstat (limited to 'tensorflow/contrib/linear_optimizer/kernels/g3doc/readme.md')
-rw-r--r-- | tensorflow/contrib/linear_optimizer/kernels/g3doc/readme.md | 2 |
1 files changed, 1 insertions, 1 deletions
diff --git a/tensorflow/contrib/linear_optimizer/kernels/g3doc/readme.md b/tensorflow/contrib/linear_optimizer/kernels/g3doc/readme.md index f5fc77b9c1..a4f5086dde 100644 --- a/tensorflow/contrib/linear_optimizer/kernels/g3doc/readme.md +++ b/tensorflow/contrib/linear_optimizer/kernels/g3doc/readme.md @@ -159,7 +159,7 @@ expected. On criteo dataset, the usual Newton method goes out of range for a small (but non negligible) fraction of the examples. The returned dual in these cases will -be $$0$$ or $$\pm 1$$. The modified Newton algorihm always find the true zero +be $$0$$ or $$\pm 1$$. The modified Newton algorithm always find the true zero and achieves a better log loss. The blue lines represent the modified Newton (evaluation and training) and the |