Introduction to Machine Learning - Introduction · 2020-04-12 · Introduction to Machine Learning...

Introduction to Machine Learning

Introduction

Varun ChandolaComputer Science & Engineering

State University of New York at BuffaloBuffalo, NY, USA

chandola@buffalo.edu

Chandola@UB CSE 474/574 1 / 16

Outline

Making Intelligent Machines

Human Learning

Definition of Machine LearningWhy Machine Learning?

Learning from (Past) Data

Overview of ML

Supervised Learning

Chandola@UB CSE 474/574 2 / 16

Making Machines Intelligent

The world is woven from billions of lives, every strand

crossing every other. What we call premonition is just

movement of the web. If you couldattenuate to every strand ofquivering data, the futurewould be entirely calculable, asinevitable as mathematics.

Sherlock Holmes, 2017

Chandola@UB CSE 474/574 3 / 16

What makes an artificial machine intelligent?

1. Talk. See. Hear.I Natural Language Processing, Computer Vision, Speech Recognition

2. Store. Access. Represent. (Knowledge)I Ontologies. Semantic Networks. Information Retrieval.

3. Reason.I Mathematical Logic. Bayesian Inference.

4. Learn.I Improve with Experience

I Machine Learning

Chandola@UB CSE 474/574 4 / 16

What makes an artificial machine intelligent?

1. Talk. See. Hear.I Natural Language Processing, Computer Vision, Speech Recognition

2. Store. Access. Represent. (Knowledge)I Ontologies. Semantic Networks. Information Retrieval.

3. Reason.I Mathematical Logic. Bayesian Inference.

4. Learn.I Improve with Experience

I Machine Learning

Chandola@UB CSE 474/574 4 / 16

Human Learning?

I What do we learn?I Concepts (this is a chair, that is not a chair)I Distinguishing concepts (this is a chair, that is a table)I Other things (language, juggling, using a remote)

I How do we learn?

1. Teaching (Passive).2. Experience (Active).

2.1 Examples.2.2 Queries.2.3 Experimentation.

Chandola@UB CSE 474/574 5 / 16

What is Machine Learning?

I Computers learn without being explicitly programmed.I Arthur Samuel (1959)

I A computer program learns from experience E with respect to sometask T, if its performance P while performing task T improves overE.I Tom Mitchell (1989)

Chandola@UB CSE 474/574 6 / 16

Why Machine Learning?

I Machines that know everything fromthe beginning?I Too bulky. Creator already knows

everything. Fails with newexperiences.

I Machines that learn?I Compact. Learn what is necessary.I Adapt.I Assumption: Future experiences are

not too different from pastexperiences.

I Have (structural) relationship.

Chandola@UB CSE 474/574 7 / 16

Learning from (Past) Data

Deductive LogicI All birds can fly

I Dodo is a bird

I ⇒ Dodo can fly

Inductive LogicI A stingray can swim

I Stingray is a fish

I ⇒ All fish can swim

Core TenetI Deduce Induce from past

I Generalize for future

Chandola@UB CSE 474/574 8 / 16

Learning from (Past) Data

Deductive LogicI All birds can fly

I Dodo is a bird

I ⇒ Dodo can fly

Inductive LogicI A stingray can swim

I Stingray is a fish

I ⇒ All fish can swim

Core TenetI Deduce Induce from past

I Generalize for future

Chandola@UB CSE 474/574 8 / 16

Understanding the Machine Learning Landscape

I What do you want the ML algorithm to do?

I How do you want to do it?

Chandola@UB CSE 474/574 9 / 16

Machine Learning Problems

Supervised LearningI Given a finite set

of x’s andcorresponding y ’s,learn f ()

I Infer y for a new x

I y - continuous(regression)

I y - discrete(classification)

UnsupervisedLearningI Given only x’s,

infer structure indataI hidden (latent)

relationshipsamong theobjects

I e.g., clustering,embedding,dimensionalityreduction, etc.

ReinforcementLearning*I Find the best

mapping ofsituations toactions tomaximize anumerical reward

I Agent learns tobehave in anenvironment

Chandola@UB CSE 474/574 10 / 16

Machine Learning meet Mother Nature

I Data: A collection of data objectsI A representation of the object of interest or

the state of the target system, and/orI A label or a target value or a low-dimensional

representation or an embedding or an actionassociated with the object/state

I Mother nature generates this data using anunknown generative process (secret recipe)

I ML problem - given part of the data, infer theother part

Chandola@UB CSE 474/574 11 / 16

Supervised Learning

https://quickdraw.withgoogle.com/

Chandola@UB CSE 474/574 12 / 16

More Examples

PrimaryI Given an individual’s data (criminal and otherwise), predict risk of

recidivism

I Given an email, determine if it is spam or normal

I Given an image, identify the object

I For a given day, predict the number of travelers that will passthrough a subway station

SecondaryI Color a black and white image

I Translate English text to French

Chandola@UB CSE 474/574 13 / 16

Unsupervised Learning

I Learn hidden structure in the data

I No assumption of labels

I Examples

1. Clustering (customers of Wegmans, set of magazine articles)2. Embedding any data into a metric space (<d) (text, tweets, disease

codes)3. Dimensionality reduction4. Factor analysis5. Dictionary learning

Chandola@UB CSE 474/574 14 / 16

Reinforcement Learning

I An agent, operating in a changing environment, finds the bestsequence of actions that maximize an end goal

I Key concepts: policy, reward, end-value, environment

I Examples

1. Learning to play a game (Breakout, AlphaGo)2. A robotic vaccum cleaner figures out the best time to recharge3. Almost all robotic tasks4. Traffic light control

Chandola@UB CSE 474/574 15 / 16

References

Chandola@UB CSE 474/574 16 / 16