Slides from my presentation on Introduction to Deep Learning on Azure presented at Global Azure Bootcamp 2018.

1. INTRODUCTION TO DEEP LEARNING ON
AZURE
SHASHI JEEVAN M P
@shashijeevan
https://shashijeevan.com

2. DEFINITION
• Artificial Intelligence
• Machine Learning
• Deep Learning

5. MACHINE LEARNING DEFINITION
• Machine Learning: Field of study that gives computers the
ability to learn without being explicitly programmed.
~Arthur Samuel, 1959
• Well posed Learning Problem: A computer program is said to
learn from experience E with respect to some task T and some
performance measure P, if its performance on T, as measured
by P, improves with experience E.
~Tom Mitchell, 1998

6. MACHINE LEARNING TYPES
• Supervised Learning – Naïve Bayes, SVM, Artificial Neural Nets,
Random Forest
• Unsupervised Learning – K Means Clustering
• Reinforcement Learning – Model Free Learning, MDP, Q
Learning
• Semi Supervised Learning – GAN (New)

7. PREDICTION
• Regression
• Classification

8. HOW TO PREDICT HOUSE PRICE?
• Identify the features and price
• Train using the samples with known price
• Use the new house data and predict price

9. LINEAR REGRESSION
• Best fit line
Y = m * X + B
Minimize Errors using Least Squares method

10. APPLICATIONS
• Netflix
• Uber
https://eng.uber.com/michelangelo/
• Amazon shopping

11. Machine Learning Reference Architecture

12. NEURAL NETWORKS
• Huge number of features (Hundreds)
• Identify the features automatically

14. DEEP NEURAL NETWORKS
• Specialized Neural Networks
• Lots of Hidden layers (Typically Hundreds)
• Uses latest Techniques
• ReLU activation
• Residual Networks
• Specific architectures to solve specific problems

17. NETWORK ARCHITECTURES
• https://becominghuman.ai/cheat-sheets-for-ai-neural-
networks-machine-learning-deep-learning-big-data-
678c51b4b463

22. MICROSOFT ARCHITECTURE
• Residual Networks
• 152 Layers
• De-facto Standard for image classification with CNN

24. JARGON
• Loss Function
• Gradient Descent
• Back propagation

25. TYPES OF APPLICATIONS
• Classification
• Regression

26. POPULAR USES
• Self driving cars
• Text Translations
• Automatic colorization of B&W images
• Sentiment Analysis

29. PIPELINE
• Build Model
• Train
• Inference for Prediction

30. BUILDING MODEL
• Layers are organized
• Activation function –ReLU, Sigmoid, Softmax
• Loss function
• Optimizer

31. TRAINING
• Training Set/Testing Set
• Batch
• Weights adjustment
• Epoch
• Loss
• Accuracy

32. INFERENCE
• Saved model is used
• Untrained Input is processed and a prediction is made

33. AZURE
• Data Science VM
• Deep Learning VM
• GPU support (NC- Series)
• Use Azure Trial (need to convert to Paid account)
• Trial account has 4 cores
• Deep Learning VM requires minimum 6 cores
• Specific Regions
• East US, East US 2, North Central US, South Central US and West US 2

34. AZURE SUPPORTED
• Nvidia GPU
• GPU version of Deep Learning Frameworks
• Jupyter Hub
• X2Go

35. SETUP ON AZURE
• Create Deep Learning VM Ubuntu
• Setup Auto shutdown
• Connect using SSH with the hostname (IP Address is reset on
restart)
• Build Model
• Train
• Infer

36. TYPICAL USES
• Ready to use for development
• Consistent setup for a team
• Use it for a temporary training tasks

37. DEMO
• Training
• MNIST dataset
• Keras with Microsoft CNTK backend
• Inference
• Use trained model
• Predict the digit in Jupyter
• Write and Predict

38. DEMO CODE LINKS
• Jupyter Notebooks for Training and Prediction
https://github.com/shashijeevan/keras_mnist_notebooks
• Python Flask App for generating digits and predict
https://github.com/shashijeevan/mnist-draw

40. NEXT STEPS
• ONNX – Model Exchange Standard
• Windows ML – Inference Built into Windows
• Visual Studio Tools for AI

41. RESOURCES
• https://academy.microsoft.com/en-us/professional-
program/tracks/artificial-intelligence/
• https://www.analyticsvidhya.com
• https://www.coursera.org/
• https://machinelearningmastery.com/handwritten-digit-
recognition-using-convolutional-neural-networks-python-
keras/