2015-3-31

转载于 2015-03-31 23:11:00 发布 · 60 阅读

0 ·

CC 4.0 BY-SA版权

原文链接：http://www.cnblogs.com/peng-ge/p/4382434.html

文章标签：

#人工智能

作者在Interspeech2015中尝试使用BLSTM进行分类，但效果未达预期。为提升准确率，考虑将底层BLSTM层替换为CNN进行重新测试。同时阅读了DRAW和RAM两篇论文，虽未获得深入见解，但认识到这些工作均涉及注意力机制的学习，且RAM是POMDP的特定案例，而DRAW则侧重于网络上的变分推断。

Long time no blog.

I worked on Interspeech 2015, but failed. The classification accuracy is not as good as excepted. I will change the lower BLSTM layer to CNN to do another test.

I read DRAW yesterday and RAM today. I do not have any insights on these works. I need to learn a lot of prior knowledge.

RAM is a specific case of POMDP and is more of reinforcement learning. DRAW is more of variational inference on networks. But they all learn some attention mechinism.

All these models are important. Memory networks need to label which sentence is useful. Probably there is a reinforement learning way.

Memory, attention, reinforcement learning exists on this.

Fightting.

转载于:https://www.cnblogs.com/peng-ge/p/4382434.html