SSIST by-talk Day 1

The first impression of ShanghaiTech University is different from other universities I’ve visited.

Cheetah Mobile AI Lab

Speaker is a young man who quit UIUC PHD conducted by MaYi and made a startup evaluated as almost 200 million yuan.

AI in product news republic for understanding content

• 500 thousand articles
• Convert content to 512 dimension vector representation by DNN plus TF-IDF and other basic vectors
• Use LSH to find the similar articles.
• LSH can’t leverage the data distribution and all article is treated as the same
• CVRP 2016 deepbit

HUAWEI

Speaker Hua.Huang, director of Mobile Broadband Network Research Department.

• Introduce the reinforcement learning (Q-learning) in network traffic optimization.

360

Speaker Shuicheng Yan, 360 VP.

Two target of AI:

• Pursue extreme score (Research)
• Pursue extreme user-experience (Industry)

Dual-path network: Residual network + densely connected network

Apply AI model need to concern memory, time performance and computation ability.

• Add a 11 convolution layer to multiply the normal nn layer to select features.
• RELU only use 60%, others are almost 0 value, no effect on the next layer
• Also adjust the model complexity can work

SenseTime

Li Xu, CEO of SenseTime. A funny person.

Panel

Extraordinary candidates:

• Confident and try
• Own different opinion and view
• Insist his own view
• Do something extreme in some aspect
• Willing to learn

Conference notes from SSIST 2017 Day 1.