Deep Learning

It’s a beautiful Friday night in Kirkland, WA. A Mandarin speaking meetup was hosted by the SeattleStartup on the subject of “Deep Learning” at 7:30pm. On my way there, I wondered who’s going to show up and why was Mandarin being used. By 7:40pm, to my surprise, there were 70+ young Chinese professionals in the room; a good number of Microsoft employees mixed with a good number of Amazon employees.

Deeplearning.net summarizes Deep Learning as follows: “Deep learning is a new area of Machine Learning research, which has been introduced with the objective of moving Marching Learning closer to one of its original goals: Artificial Intelligence.

This meetup agenda was as follows:

  1. Dr. Dong Yu, a principle researcher at Microsoft Research, Deep Learning Talk
  2. Panel Discussion about the Deep Learning Application with two AI companies:
  • Kitt AI founder Xuchen Yao and Guoguo Chen. Kitt AI is backed Paul Allen’s AI2, Amazon Alexa Fund and Madrona Venture.
  • Orbeus CEO Yi Li and her team. Orbeus was an AI startup from Silicon Valley in Image Recognition. It was acquired by Amazon and the company has recently moved to Seattle.

Dr. Yu introduced the basic concept of Deep Learning and described the key models such as Deep Neural Networks, Convolutional Neural Networks and Long Short-term Memory Recurrent Networks. He also illustrated the core design principles of Deep Learning models when introducing other new models.

He explained that Deep Learning is feasible only with the computer power and big data available today. It’s actually a rebranding and extension of neural networks. Among the three Deep Learning definitions he gave, the shortest one was “Any system that involves more than one layer of nonlinear processing”

He stated that Deep Learning’s essentials are:

  • Learn complicated feature representation through many layers of nonlinear processing
  • Learn representation automatically and jointly with classification (or whatever) tasks (end-to-end optimization)
  • Key: design the model structure and training criterion.

One of the examples given was AlphaGo; a computer program developed by Google DeepMind in London to play a board game.

AlphaGo’s deep learning is as follows

  • Supervised learning on expert games
  • Reinforcement learning; improve through self-play
  • Build a strategy network and a value network
  • Monte Carlo Tree Search (MCTS) to determine moves through real-time play

Dr. Yu continued to explain Artificial General Intelligence (AGI)

  • The intelligence of a machine that could successfully perform any intellectual task that a human being can (Wikipedia)
  • An emerging field aiming at the building of “thinking machines”; i.e.: General purpose systems with intelligence comparable to that of the human mind (agi-society.org)
  • The general-purpose mechanisms and learning principles that allow machines to explore the world, form connects and clusters, develop and validate theories, learn and generalize from a small number of examples and reason and plan with uncertainty.

The Panel Discussion was quite interesting. In order to learn these two startup’s journey and products, the audience was very actively asking questions and obtaining information. Since these are Chinese startups, it seems using native language Mandarin is a natural way to network and brain storm. That answered my question: why Mandarin.

Personally, I was very happy to have the opportunity to learn about Deep Learning and was impressed by the positive energy brought by these young professionals. I see entrepreneurship emerging among the corporate employees and that’s a good sign. My next blog will explain my perspective about the entrepreneurial culture in corporation. Stay tuned!

[This was originally posted on Linkedin on 5-3-16]