Day 1- Logistics and Background

Posted on May 29, 2017 by Govind Gopakumar

Lecture slides in pdf form

Introduction

Rules and Administration

Logistics

Course Outline

Course goals and outcomes

Origins of Machine Learning / AI

All of these come together and now, we know them by the general terms Artificial Intelligence, Machine Learning, Data Science

Machine Learning : the present

Teaching Computers to think

“When we write programs that ‘learn’, it turns out that we do and they don’t.” - Alan Perils

All of these are examples of Machine learning in action!

Common element - Some form of data, expectation of some trend

What constitutes trend and data?

How would you model this information? How would you make predictions?

Detailed system - autonomous drones

Involved systems :

All these involve some form or the other of Machine Learning!

Formal description of Machine Learning

Questions that we need to ask :

  1. How do you get this data? (Feature engineering)
  2. How do you choose an appropriate model? (Model selection)
  3. How do you train this efficiently? (Optimization and Learning)
  4. How do you use this model? (Inference)
  5. Why should this work? (Learning theory!)

Supervised Learning

Predict an outcome, a value, or a class. What do we have?

What is our goal?

Supervised Learning : Examples

Classification

Regression

Labelling

Unsupervised Learning

Find out patterns, modify the data automatically. What do we have?

What is our goal?

Unsupervised Learning : Examples

Clustering

Density estimation

Dimensionality reduction

The Mathematics

For ease of use, we’ll always refer consider two phases :

Training a model

What steps do we take to train a model?

This is where words like “gradient descent” and “optimization” come in.

We need to find a right setting of parameters that do well by our measure.

Toy example - 2D Classification

Input Data : points in 2D space, with labels

Model selected : Straight line (unknown slope, offset)

Train time - Find proper splitting

Test time - Assign label to a new point

Machine Learning : The Future

Current applications

Tentative syllabus and breakdown

Lecture takeaways

Code

Policy

No code for this lecture, but further lectures will have associated code. Will be provided in the form of an IPython notebook, for ease of use and reproducibility.

References

Optional reading

Next class

Overview of mathematics - Linear Algebra, Probability

Next class Overview

Linear Algebra

What you must be comfortable with :

Probability

Functions - Shapes, Optima

Shape of a function

Optima finding