The purpose of these sessions is to provide a quick high-level introduction to Machine learning, We will cover Neural Networks, Convolution Neural Networks, Transfer Learning and Model Ensemble techniques. We will build a system that can recognize a chart type from its image and we will walk through many cycles to take its accuracy from 65% to 99% with only 40 to 60 samples of training for each chart type. Python Notebooks will be provided for each step so you can follow along and experiment with the code yourself.
-
Session 1: Introduction to Machine Learning, building examples showing how to use TensorFlow to approximate a simple linear function, a quadratic function, and XOR.
-
Session 2: How to recognize the type of chart from an image that includes the chart. We will cover How to load images and prepare them to be suitable for use in Neural Network training.
-
Session 3: Improve the accuracy of the chart recognizer done in session 2, by using Convolutional Neural Network. We will also show how this reduces the model size dramatically and speeds up training.
-
Session 4: Using transfer learning techniques to improve the accuracy of the chart recognizer we built-in session 3, and make the training process go faster.
-
Session 5: Combine different pre-trained models to improve the accuracy of the model without adding new training samples.
-
Session 6: Introduction to time series forecasting with deeplearning
-
Session 7 : Introduction to Generative Adversarial Networks (GAN)
-
Session 8: How the model see your image through the different layers. (code only, write-up coming soon)
-
Session 9: Wrap the model by a rest service and create a docker image to allow deploying it easily. (code only, write-up coming soon)