“Zhiyuan Thang 15-06-29”版本间的差异
来自cslt Wiki
（以“ Last few weeks (lazy): Presently got the conclusion that randoming the weights ahead of softmax of DNN before MPE training may be not helpful. Using language vec...”为内容创建页面） |
|||
| 第2行： | 第2行： | ||
Last few weeks (lazy): | Last few weeks (lazy): | ||
| − | + | presently got the conclusion that randoming the weights ahead of softmax of DNN before MPE training may be not helpful; | |
| − | + | using language vector to pre-train the hidden layers of DNN gives at least a little improvement, especially 3 hidden layers is pre-trained when DNN is of 4 hidden layers; | |
| + | |||
| + | have not realized leaky Rectifier and temporal Rectifier, as CUDA programing may be needed; | ||
| + | |||
| + | paper reading. | ||
| − | |||
This week: | This week: | ||
| − | realize the two kinds of Rectifier. | + | realize the two kinds of Rectifier; |
| + | |||
| + | paper reading. | ||
2015年6月30日 (二) 14:36的版本
Last few weeks (lazy):
presently got the conclusion that randoming the weights ahead of softmax of DNN before MPE training may be not helpful;
using language vector to pre-train the hidden layers of DNN gives at least a little improvement, especially 3 hidden layers is pre-trained when DNN is of 4 hidden layers;
have not realized leaky Rectifier and temporal Rectifier, as CUDA programing may be needed;
paper reading.
This week:
realize the two kinds of Rectifier;
paper reading.
