"Deep Learning" AI Frontier Conference

Starting with the opening keynote by Jeff Dean, the lead in Google Brain, the 2017 AI Frontier Conference is full of great speeches with technical leads from Amazon, Google, Facebook, Baidu, Microsoft, Tesla etc.

A few key things I have learned from these speeches.

(1) Self-driving car progress and challenge

There are many kind of sensors and because of the AI technology advance, even an "gesture wave" from a bike rider can be detected, which is amazing.

Both Google and Baidu has demonstrated cool real world driving videos.

The current challenges from leading self-driving companies:

• Google - Safety enhancement
• Tesla - Cross-domain machine learning
• Baidu - Integration with car and power battery

From the conference impression, in 2-3 years, the technology should be truly mature.

Two message I got from the conference for the benefits of the self-driving car

• Safety improvement and reduce car accident (US has 1M+ car death number per year)
• Reduce a lot on parking and parking introduced traffic (30%)

(2) Chatbot

Microsoft's Li Deng, Baidu AI lab's Adam Coates and Amazon's Nikko Strom (Alexa) have all pointed to a rosy future of speech recognition.

The user experience of voice driving Apps using Chatbot can change the world because of these factors:

• Speech is 3X faster than typing
• Speech has less error chance than typing
• Voice recognition is now really working
• Chatbot can be very flexible to do personalization
• Alexa can detect who is speaking even in a party environment

Some people probably still are thinking: I can type and why I should speak. But maybe need to think again.

(3) Jeff Dean's 4 example queries of the future

Image recognition has become so good now which helps self-driving and now next is to move to do video recognition. (Youtube is releasing 1M+ video dataset to public to research)

Jeff pointed out these future questions:

1. Which of these eye images shows symptoms of diabetic retinopathy?
2. Please fetch me a cup of tea from the kitchen
3. describe this video in Spanish
4. Find me documents related to reinforcement learning for robotics and summarize them in German

(4) AI deep learning frameworks and computing engine

With many deep learning framework in the market, which one to choose?

But Google's tensorflow is worth to mentioned:

In only about 1 year, it has reached 1M+ downloads and 5000+ Tensorflow related repos in Github.

TensorFlow's Rajat Monga, Facebook's Yangqing Jia and Soumith Chintale and Amazon's Alexa Smola all gave great presentations to the framework often used in the industry: TensorFlow(google), Caffe(Facebook), Torch(Facebook), MxNet(Amazon)

There are many reasons for considering a framework including performance, ease to use, easy to port, great community, easy to deploy. H2O Deep Water and Amazon DL image ami includes all major open source frameworks for customers to choose.

Consideration on portability and distribution deep learning are also important.

One of the work done by Professor Hai Tao on image recognition (can count the heads of the supermarket by a camera) showed some performance difference on FPGA and GPU is interesting. Some use cases FPGA is much faster with some limitation comparing to GPU.

(5) Job section can be automated

Around 50-60% jobs can be automated based on prediction presented by Mckinsey.

These sectors shall be impacted

Summary

In general, the first day experience is great. The organizers have done great job to arrange. I have not mentioned the fun use case demoed by Google - farmers using TensorFlow to sort cucumber by sizes, really helpful.

Look forward to a great year of AI, 2017! One area I like to explore more is that how to make ML to improve daily life like you and me to make decisions and make our life truly productive and fun!