Quality Alchemist (品質煉金術師): Google AI Teacher Camp

2019年1月18日星期五

Google AI Teacher Camp

The Google AI Teacher Camp is for university academic/research members The event covered concepts and application of Tensorflow. They would be sharing ways to learn about core machine learning concepts, develop and hone participants’ Machine Learning (ML) skills via Tensorflow, and deep dive into ML on mobiles. The Google AI Teacher Camp was held on 18^th Jan 2018. I took a photo in the entry of Google office.

I met my colleague in SEEM Dept. CityU – Dr. Zhang Qingpeng.

Mr. Laurence Moroney was the trainer of this Google AI Teacher Camp.

There were three topic tonight and they were “Mobile Machine Learning – AI in the palm of your hand”, “TensorFlow for Javascript” and “TensorFlow: Autograph for Easier Code”.

Firstly, Mr. Laurence Moroney showed ML located in the technology trigger with sharp growth trend.

Then he explained the different scope of AI, ML and DL. (Remark: the subtitle in each diagram was employed Google instant translator.)

The different between Traditional Programming and Machine Learning was mentioned. For Traditional Programming, inputs are Rules and Data, and then output is answers. For Machine Learning, inputs are Answers and Data, and then output is Rules. After that he also mentioned about Model that its input is Data and output is Predictions.

He use activity detection as example to explain different sport.

After that Mr. Moroney demonstrated Tensorflow using Python to find out the Predicted Model (equation) by data X and Y.

He also used Fashion-MNIST (which is a dataset of Zalando's article images—consisting of a training set of 60,000 examples and a test set of 10,000 examples.) to import into tensorflow for training.

The following diagram showed TensorFlow and TensorFlow Lite in Training Phase and Inference Phase, respectively.

And then he briefed to transfer the model from workstation to mobile using TensorFlow Lite Format.

A video “TensorFlow: an ML platform for solving impactful and challenging problems” was demonstrated.

The second topic was TensorFlow for JavaScript. Firstly, Mr. Moroney demonstrated the ML structure included Data, Features, Hidden Layers and Output. In-browser ML needed no drivers/No install, but had interactive, sensors and data stays on the client.

Then Mr. Moroney introduced TensorFlow.js and its API structure. (TensorFlow.js is an open source WebGL-accelerated JavaScript library for machine intelligence. It brings highly performant machine learning building blocks to your fingertips, allowing you to train neural networks in a browser or run pre-trained models in inference mode.) And then he showed the Core API to fit a polynomial codes and Layers API for Speech Command Recognition.

The coding and layer of NN was demonstrated and explained.

After that he used Pacman game using webcam to control the movement of the Pacman.

Mr. Moroney also compared the speech in deferent hardware running TensorFlow.

The last session was TensorFlow Autograph. Mr. Moroney demonstrated Control flow in graphs.

TensorFlow Eager was introduced at https://www.tensorflow.org/guide/eager . Eager mode and Graph mode are complementary.

AutoGraph was then introduced that it helped to write complicated graph code using normal Python. Details at https://www.tensorflow.org/guide/autograph . The AutoGraph features were described as following diagram.

Finally, Mr. Moroney mentioned the TensorFlow 2.0 would be coming. Moreover the training and deployment structure was described. At the end, Machine Learning Crash Course was introduced at https://developers.google.com/machine-learning/crash-course/ .