Practical Artificial Intelligence and Machine Learning arturo.servin_at_gmail.com http://arturo.servin.googlepages.com/

arturo.servin_at_gmail.com

http://arturo.servin.googlepages.com/

About this presentation Some theory on AI and ML Some practical ideas and simple how to What's out there using AI Resources, Kits and Data

Some theory on AI and ML

Some practical ideas and simple how to

What's out there using AI

Resources, Kits and Data

Artificial Intelligence Machine Learning Natural Language Processing Knowledge representation Plannning Multi-Agent Systems and some other stuff depending of the author of the book

Machine Learning

Natural Language Processing

Knowledge representation

Plannning

Multi-Agent Systems

and some other stuff depending of the author of the book

Machine Learning A program is learning when it executes a task T and acquires experience E and the measured performance P of T improves with experience E (T. Mitchell, Machine Learning, 1997)

A program is learning when it executes a task T and acquires experience E and the measured performance P of T improves with experience E (T. Mitchell, Machine Learning, 1997)

Machine Learning Flavours Supervised Learning Programs learn a concept/hypothesis by means of labeled examples Examples: Artificial Neural Networks, Bayesian Methods, Decision Trees Unsupervised Learning Programs learn to categorise unlabelled examples Examples: Non-negative matrix factorization and self-organising maps

Supervised Learning

Programs learn a concept/hypothesis by means of labeled examples

Examples: Artificial Neural Networks, Bayesian Methods, Decision Trees

Unsupervised Learning

Programs learn to categorise unlabelled examples

Examples: Non-negative matrix factorization and self-organising maps

More flavours Reinforcement Learning Programs learn interacting with the environment, the execution of actions and observing the feedback in the form of + or – rewards Examples: SARSA, Q-Learning

Reinforcement Learning

Programs learn interacting with the environment, the execution of actions and observing the feedback in the form of + or – rewards

Examples: SARSA, Q-Learning

Training Examples Continuous Discrete Inputs know as Vectors or Features Example in Wine Classification: Alcohol level, Malic acid, Ash, Alcalinity of ash, etc.

Continuous

Discrete

Inputs know as Vectors or Features

Example in Wine Classification: Alcohol level, Malic acid, Ash, Alcalinity of ash, etc.

Linear and Non-linear feature relations source: Oracle Data Mining Concepts

Linear and Non-linear feature relations

source: Oracle Data Mining Concepts

More complex feature relations

Decision Trees Easy to understand and to interpret Hierarchical structure They use Entropy and Gini impurity to create groups Disadvantage: It's an off-line method Examples: ID3, C4.5 Source: http://www2.cs.uregina.ca/~dbd/cs831/notes/ml/dtrees/c4.5

Easy to understand and to interpret

Hierarchical structure

They use Entropy and Gini impurity to create groups

Disadvantage: It's an off-line method

Examples: ID3, C4.5

Decision Trees, example Create .names and .data files with training data Generate tree and rules (c4.5 -f <file> and c4.5rules -f <file>) Outlook Temperature Humidity Windy Play or Don't Play Sunny 80 90 true Don't Play Overcast 83 78 false Play Rain 70 96 false Play Categorize new data (consult,consultr). Use GPS/Geocoding, Google Maps and Yahoo Weather APIs to enhance aservin@turin:~/Projects/C45: consult -f golf C4.5 [release 8] decision tree interpreter Sat Jan 17 00:05:16 2009 ------------------------------------------ outlook: sunny humidity: 80 Decision: Don't Play CF = 1.00 [ 0.63 - 1.00 ]

Create .names and .data files with training data

Generate tree and rules (c4.5 -f <file> and c4.5rules -f <file>)

Outlook Temperature Humidity Windy Play or Don't Play

Sunny 80 90 true Don't Play

Overcast 83 78 false Play

Rain 70 96 false Play

Categorize new data (consult,consultr). Use GPS/Geocoding, Google Maps and Yahoo Weather APIs to enhance

aservin@turin:~/Projects/C45: consult -f golf

C4.5 [release 8] decision tree interpreter Sat Jan 17 00:05:16 2009

------------------------------------------

outlook: sunny

humidity: 80

Decision:

Don't Play CF = 1.00 [ 0.63 - 1.00 ]

Bayesian Classifiers Bayes Theorem: P(h|D) = P(D|h) P(h) P(D) P(h) = prior probability of hypothesis h P(D) = prior probability of training data D P(h|D) = probability of h given D P(D|h) = probability of D given h Naive Bayes Classifier, Fisher Classifier Commonly used in SPAM filters

Bayes Theorem: P(h|D) = P(D|h) P(h) P(D)

P(h) = prior probability of hypothesis h

P(D) = prior probability of training data D

P(h|D) = probability of h given D

P(D|h) = probability of D given h

Naive Bayes Classifier, Fisher Classifier

Commonly used in SPAM filters

Classifying your RSS feeds Use the unofficial Google Reader API http://blog.gpowered.net/2007/08/google-reader-api-functions.html Some Python Code (Programming Collective Intelligence, Chapter 6) Tag interesting and non-interesting items Train using Naive-bayes or Fisher classsifier >> cl.train('Google changes favicon','bad') >> cl.train('SearchWiki: make search your own','good') New items are tagged as interesting or not >> cl.classify('Ignite Leeds Today') Good You can re-train online Add more features, try with e-mail

Use the unofficial Google Reader API http://blog.gpowered.net/2007/08/google-reader-api-functions.html

Some Python Code (Programming Collective Intelligence, Chapter 6)

Tag interesting and non-interesting items

Train using Naive-bayes or Fisher classsifier

>> cl.train('Google changes favicon','bad')

>> cl.train('SearchWiki: make search your own','good')

New items are tagged as interesting or not

>> cl.classify('Ignite Leeds Today')

Good

You can re-train online

Add more features, try with e-mail

Finding Similarity Euclidean Distance, Pearson Correlation Score, Manhattan Document Clustering Price Prediction Item similarity k-Nearest Neighbors, k-means, Hierarchical Clustering, Support-Vector Machines, Kernel Methods

Euclidean Distance, Pearson Correlation Score, Manhattan

Document Clustering

Price Prediction

Item similarity

k-Nearest Neighbors, k-means, Hierarchical Clustering, Support-Vector Machines, Kernel Methods

Similar items Source: http://home.dei.polimi.it/matteucc/Clustering/tutorial_html

Artificial Neural Networks Mathematical/Computational model based on biological neural networks Many types. The most common use Backpropagation algorithm for training and Feedforward algorithm to get results/training

Mathematical/Computational model based on biological neural networks

Many types. The most common use Backpropagation algorithm for training and Feedforward algorithm to get results/training

Artificial Neural Networks Input is high-dimensional, discrete or real-valued (e.g. raw sensor input) Output is discrete or real valued Output is a vector of values Perceptron, linear Sigmoid, non-linear and multi-layer

Input is high-dimensional, discrete or real-valued (e.g. raw sensor input)

Output is discrete or real valued

Output is a vector of values

Perceptron, linear

Sigmoid, non-linear and multi-layer

Example, finding the best price Create training data using Amazon/Ebay API Laptop prices. Use price, screen size as features Use a ANN, i.e. Fast Artificial Neural Network (FANN) struct fann *ann = fann_create_standard(num_layers, num_input, num_neurons_hidden, num_output); #C++ ann = fann.create(connection_rate, (num_input, num_neurons_hidden, num_output)) #Python $ann = fann_create(array(2, 4, 1),1.0,0.7); // PHP You can also try k-Nearest Neighbours Try it!

Create training data using Amazon/Ebay API

Laptop prices. Use price, screen size as features

Use a ANN, i.e. Fast Artificial Neural Network (FANN)

struct fann *ann = fann_create_standard(num_layers, num_input, num_neurons_hidden, num_output); #C++

ann = fann.create(connection_rate, (num_input, num_neurons_hidden, num_output)) #Python

$ann = fann_create(array(2, 4, 1),1.0,0.7); // PHP

You can also try k-Nearest Neighbours

Try it!

Resources 1 Books Practical Artificial Intelligence Programming in Java, Mark Watson http://www.markwatson.com/opencontent/ (There is a Ruby one as well) Programming Collective Intelligence, Toby Segaran; O'Reilly Artificial Intelligence: A Modern Approach, S. Russell, P. Norvig, J. Canny; Prentice Hall, Machine Learning, Tom Mitchell; MIT Press Online Stuff ML course in Stanford http://www.stanford.edu/class/cs229/materials.html Statistical ML http://bengio .abracadoudou.com/lectures/

Books

Practical Artificial Intelligence Programming in Java, Mark Watson http://www.markwatson.com/opencontent/ (There is a Ruby one as well)

Programming Collective Intelligence, Toby Segaran; O'Reilly

Artificial Intelligence: A Modern Approach, S. Russell, P. Norvig, J. Canny; Prentice Hall,

Machine Learning, Tom Mitchell; MIT Press

Online Stuff

ML course in Stanford http://www.stanford.edu/class/cs229/materials.html

Statistical ML http://bengio .abracadoudou.com/lectures/

Resources 2 Code FANN http://leenissen.dk/fann/index.php NLP http://opennlp.sourceforge.net/ C4.5 http://www2.cs.uregina.ca/~dbd/cs831/notes/ml/dtrees/c4.5/tutorial.html ML and Java http://www.developer.com/java/other/article.php/10936_1559871_1 Data UC Irvine Machine Learning Repository http://archive.ics.uci.edu/ml/ Amazon Public Datasets http://aws.amazon.com/publicdatasets/

Code

FANN http://leenissen.dk/fann/index.php

NLP http://opennlp.sourceforge.net/

C4.5 http://www2.cs.uregina.ca/~dbd/cs831/notes/ml/dtrees/c4.5/tutorial.html

ML and Java http://www.developer.com/java/other/article.php/10936_1559871_1

Data

UC Irvine Machine Learning Repository http://archive.ics.uci.edu/ml/

Amazon Public Datasets http://aws.amazon.com/publicdatasets/

More info For questions, projects and job offers: arturo.servin \_(at)\_ gmail.com http://twitter.com/the_real_r2d2 http://arturo.servin.googlepages.com/

For questions, projects and job offers:

arturo.servin \_(at)\_ gmail.com

http://twitter.com/the_real_r2d2

http://arturo.servin.googlepages.com/