Machine Learning L1

Week 1

What Is Machine Learning?

A computer program is said to learn from an 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.

Traditional Approach to AI

Typically uses a list of hand written logical
rules

Works well if the domain is simple and well
understood

Makes extensive use of the domain
knowledge of the designer

Machine Learning Approach to AI

Typically results in models (e.g
mathematical) of the domain

Due to extensive training time it provides
better value if the domain is complex, not
well understood and has lots of data

Useful where the rules of the domain are
not very clearly defined

Four Components of ML

Assumption

Model

Interference Paradigm

Interference Engine

Assumption

What we think the world looks like.

Eg - An apple’s height and time to the ground are related

Model

A way of expressing the thought (mathematically)

Eg The relationship between an apple’s height and time to the ground is linear, quadratic ect

Interference Paradigm

A framework for matching the model to the world

Eg The difference between measured and predicted time to the ground, sort of the performance metric

Interference engine

A way of doing the matching

Eg Tweaking the model coefficients - adjusting the model to new data

Types of training (4)

Supervised

Unsupervised

Semi-supervised

Reinforcement Learning

Supervised Learning

Learn from Labeled data

Give ML algo a bunch of examples with associated labels

Classification and Regression

Classification

Goal: Predict a label or category

Output is from a fixed set of options

A type of supervised learning

Using the learnt labels, it is able to classify new instances.

Eg. Spam filter

Regression

Goal: Predict a continuous number

Output can be any value along a range

Developing a model that predicts the value of an item e.g house price given predictor number of rooms

Unsupervised Learning

Find patterns in unlabelled data

Clustering

Visualisation and dimensionality reduction

Association rule learning

Clustering

Group similar datapoints together

Eg group customers by buying behaviour

Visualisation and dimensionality reduction

Goal: Simplify complex data while keeping important information

High number of variables are hard to visualise and process

Association Rule learning

Find interesting relationships between variables

Eg people who purchase BBQ sauce and potato chips tend to buy stake

Discovers rules like:
“If A happens, B is likely to happen”

Semi Supervised

A small amount of labelled data and a large amount of unlabelled data

Small amount of labelled data used to initially train the system

System is then used to classify the unlabelled data

Benefits of semi-Supervised learning

Improved learning accuracy over unsupervised learning

but without the time and costs needed for supervised learning

Often used when you can get lots of unlabelled data from a domain but tagging them or labelling them is costly in terms of time

Eg Tagging friends and family in photos app so that the algo can predict who is who

Reinforcement Learning

Learn by trial and error with rewards and punishments

Agent interacts with environment

takes actions and gains feedback

Elements of Reinforcement Learning

Environment - Physical world in which the agent operates

State - Current Situation of the agent

Reward - Feedback from the environment

Policy - method to map agents state to actions

Value - future reward that an agent would receive by taking an action in a particular state

How they use data to learn - two types

Batch Learning - Offline Learning

Learning on the fly - Online Learning

Batch Learning

The system is trained using all available data offline. Once trained, the system is launched into production without learning again

Advantages/Disadvantages of Batch Learning

Advantage:

Problems in data are dealt with before deployment

Disadvantage:

Training can take a long time and requires lots of data

Dangers of domain overfitting

Uses a lot of computing resources

Learning on the fly/Online Learning

The system is fed with data instances in small groups called mini-batches

Advantages/Disadvantages of Online Learning

Advantages:

Each learning step is fast and cheap so the system can learn about new data on the fly as it arrives

Does not require a lot of training data

Models are adapting with time and so do not overfit to data

Disadvantages:

Prone to wrong models due to errors in data