Introduction

Overview

Teaching: 60 min
Exercises: 30 min

Questions

What is Machine Learning?

Why would you use it?

When should you use it and when should you avoid it?

Objectives

Understand the distinction between ML and traditional programming

Understand why it can be useful

Know limitations and requirements for implementation

Identify different types of ML

Why to ML?

Fundamental

Science is fundamentally data driven.


RV Investigator


Australian Square Kilometer Array Pathfinder


Energy Sector

Fast

Machine (Deep) Learning has become the poster child for fast, scalable, exascale compute.


Summit Supercomputer

Smart

Machine Learning is able to uncover insights in very complex systems without knowledge of the fundamental underlying models.


Machine Learning offers a purely data driven approach to scientific discovery.

What is ML?

FUJI or PINK LADY

Traditional models are explicitly programmed
ML algorithms are learned
Therefore data is very important
Predicition vs. understanding
Interpretability

When to ML?

Data requirements: LOTS!!! The more the better. Very generally speaking you want on the order of 10-100 samples for each feature in your data. So for example in the image processing case where we have 128x128x3 pixels we would want on the order of 500,000 -> 5,000,000 samples.
Compute requirements: In the case of deep learning the field is mostly focused on GPUs.
Desired outcomes: Supervised machine learning in particular isn’t overly useful for providing insight into data. It is mostly designed for inferencing or predictions. Unsupervsied learning on the other hand offers the opportunity to uncover patterns and relationships in data that were previously unseen.

Types of ML

Supervised (labelled data)
- Classification
- Regression
Unsupervised (unlabelled data)
- Clustering / segmentation
- Dimension reduction
Reinforcement Learning
- Environemt, state, action
Deep learning