[Adapt] [Seminar] ORDERED NEURONS: Integrating Tree Structures into RNNs

[sh]

Sorry guys, the time should be '16:30 October 23'...

----- 原始邮件 -----
发件人: "sh" <kiwisher at sjtu.edu.cn>
收件人: "adapt" <adapt at cs.sjtu.edu.cn>
发送时间: 星期二, 2019年 10 月 22日 下午 3:11:47
主题: [Adapt] [Seminar] ORDERED NEURONS: Integrating Tree Structures into RNNs

Hi Adapters,

In this paper(https://openreview.net/pdf?id=B1l6qiR5F7), the authors propose a novel RNN structure that integrates inductive bias to empower RNNs with the ability to model latent hierarchy of language. They introduce two master gates: forget master gate and input master gate, to guide information updates that are time-scale-aware. The model they propose are naturally differentiable, light-weight, and eligible for unsupervised grammar induction.

The authors conducted experiments on three different tasks to prove the general merits of their model.

Time: 16:30 October 22
Venue: SEIEE 3-414

Best,
Kiwi
______________