“Dongxu Zhang 2015-08-31”版本间的差异
来自cslt Wiki
（→Work done in this week） |
|||
| （相同用户的2个中间修订版本未显示） | |||
| 第1行： | 第1行： | ||
=== Work done in this week === | === Work done in this week === | ||
| − | * | + | * reviewed papers on bayesian graph, document classification. |
| − | * | + | * found out some interesting directions. |
(1)A more bayesian prior distribution over neural networks, that we can give constraint on a hidden layer so that the hidden | (1)A more bayesian prior distribution over neural networks, that we can give constraint on a hidden layer so that the hidden | ||
| − | + | layers follows a guassian distribution with a more reasonable mean value, which may be a direction of AAAI. | |
(2)sequential label learning, which can be a further work with Chaoyuan. | (2)sequential label learning, which can be a further work with Chaoyuan. | ||
| − | (3) | + | (3)Try a new topic model, which is similar to cbow with large window size, the code is done, still need to speed up. |
| − | (4)unbalanced autoencoder, haven't considered in details. | + | (4)unbalanced autoencoder, haven't considered in details. |
| − | * | + | (5)attention-based parser, haven't considered in details. |
| − | * | + | (6)tensor recurrent network, haven't considered in details. |
| + | * Discussed ideas with Tianyi on RS and finally chose a direction, which is a deep UV structure with content knowledge. | ||
| + | * helped Chaoyuan do the baseline reproduction. | ||
=== Plan to do next week === | === Plan to do next week === | ||
* Compare the performance with and without topic distribution constraint. Try adding constraint on different layers. | * Compare the performance with and without topic distribution constraint. Try adding constraint on different layers. | ||
2015年9月1日 (二) 11:32的最后版本
Work done in this week
- reviewed papers on bayesian graph, document classification.
- found out some interesting directions.
(1)A more bayesian prior distribution over neural networks, that we can give constraint on a hidden layer so that the hidden
layers follows a guassian distribution with a more reasonable mean value, which may be a direction of AAAI.
(2)sequential label learning, which can be a further work with Chaoyuan.
(3)Try a new topic model, which is similar to cbow with large window size, the code is done, still need to speed up.
(4)unbalanced autoencoder, haven't considered in details.
(5)attention-based parser, haven't considered in details.
(6)tensor recurrent network, haven't considered in details.
- Discussed ideas with Tianyi on RS and finally chose a direction, which is a deep UV structure with content knowledge.
- helped Chaoyuan do the baseline reproduction.
Plan to do next week
- Compare the performance with and without topic distribution constraint. Try adding constraint on different layers.
