Machine learning - HT Introduction - PDF

Please download to get full document.

View again

of 47
All materials on our website are shared by users. If you have any questions about copyright issues, please report us to resolve them. We are always happy to assist you.
Information Report
Category:

Resumes & CVs

Published:

Views: 2 | Pages: 47

Extension: PDF | Download: 0

Share
Related documents
Description
Machine learning - HT Introduction Varun Kanade University of Oxford January 20, 2016 What is machine learning? 1 Machine learning and Artificial intelligence What does intelligence entail? Reasoning,
Transcript
Machine learning - HT Introduction Varun Kanade University of Oxford January 20, 2016 What is machine learning? 1 Machine learning and Artificial intelligence What does intelligence entail? Reasoning, planning, representation, learning Courses: Intelligent Systems (MT), Knowledge Representation & Reasoning (HT) Open AI Initiative: 2 Outline History of Machine Learning This Class Some Machine Learning Applications Some Practical Concerns History of Machine Learning Statistics: Ronald Fisher Three types of iris: setosa, versicolour, virginica (1936) For each flower: sepal width (x 1), sepal length (x 2), petal width (x 3), petal length (x 4) 3 Visualize Iris Data: Setosa vs Versicolor Sepal length (cm) 4 Visualize Iris Data: Setosa vs Versicolor Sepal width (cm) 4 Visualize Iris Data: Setosa vs Versicolor Sepal width (cm) Sepal length (cm) 4 History of Machine Learning Statistics: Ronald Fisher Three types of iris: setosa, versicolour, virginica (1936) For each flower: sepal width (x 1), sepal length (x 2), petal width (x 3), petal length (x 4) Find X = w 1x 1 + w 2x 2 + w 3x 3 + w 4x 4 that maximizes D 2 /S D = 4 i=1 wi(e[xi setosa] E[xi versicolour]) µ i = E[x i] S = 4 4 i=1 j=1 wiwje[(xi µi)(xj µj)] We will see that this is basically linear regression 5 History of Machine Learning Computer Science: Alan Turing Turing Test: The Imitation Game Learning Machines (Computing Machinery and Intelligence. Mind (1950)) Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child s? If this were then subjected to an appropriate course of education one would obtain the adult brain. Quantitative computational considerations How much memory would be required? How much computational power would be required? 6 History of Machine Learning Neuroscience: Frank Rosenblatt Perceptron - neurally-inspired Simple training (learning) algorithm Built using specialized hardware 1 x 1 x 2 x 3 x 4 w 0 w 1 w 2 w 3 w 4 ϕ sign(w 0 + w 1x 1 + w 4x 4) 7 Perceptron Training Algorithm Setting Get a sequence of points (x t, y t) (where only x t is observed at first) After prediction is made y t is revealed Start with w 0 some arbitrary starting weights for the perceptron 8 Perceptron Training Algorithm Setting Get a sequence of points (x t, y t) (where only x t is observed at first) After prediction is made y t is revealed Start with w 0 some arbitrary starting weights for the perceptron Algorithm 1. Suppose w t 1 are the weights after t 1 steps 2. Predict ŷ t = sign(w t 1 x t) 3. Update: If ŷ t = y t; do nothing Else set w t = w t 1 η(1 2y t)x t 8 Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action Perceptron Training Algorithm in Action What is machine learning? Some Definitions Kevin Murphy:..., we define machine learning as a set of methods that can automatically detect patterns in data, and then use the uncovered patterns to predict future data, or to perform other kinds of decision making under uncertainty.. Tom Mitchell: A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E. Same (or similar) programs work across a range of learning tasks. (though not universally) 10 Machine Learning Intersection of computer science, statistics, neuroscience/biology, engineering, optimization etc. Statistics How much data is needed? When can we be confident in our predictions? Computer Science Design algorithms for automated pattern discovery. How fast do these run? How much computational power is needed? 11 Outline History of Machine Learning This Class Some Machine Learning Applications Some Practical Concerns About this course Pre-requisites: Basic linear algebra, calculus, probability, algorithms, programming. Mathematical foundations of ML; not computational or statistical learning theory Regression, support vector machines, neural networks, deep learning, clustering [video] Conceptual programming assignments; not scaling to real-world systems 12 About this course Discussion forum on Piazza (link on webpage) Classes in Weeks 3-7 (Mon, Wed, Fri - 6 groups) Practicals in Weeks 2-8 (Tue, Thu - 2 groups) Final examination over easter break Office Hours: Tue 15:30-16:30 (449 Wolfson Building) 13 Outline History of Machine Learning This Class Some Machine Learning Applications Some Practical Concerns Application: Boston Housing Dataset Real attributes Crime rate per capita Non-retail business fraction Nitric Oxide concentration Age of house Floor area Distance to city centre Predict house cost Integer attributes Number of rooms Categorical attributes On the Charles river? Index of highway access (1-5) Source: UCI repository 14 Application: Breast Cancer Integer attributes Clump thickness Uniformity of cell size Uniformity of cell shape Marginal adhesion Single epithelial cell size Bare nuclei Bland Chromatin Normal nucleoli Mitoses Predict: Benign vs Malignant Source: UCI repository 15 Application: Object Detection and Localization 200-basic level categories Dataset contains over 400,000 images Imagenet competition ( ) 16 Application: Object Detection and Localization Source: DeepLearning.net (top); Brain-Maps.com (bottom) 17 Application: Object Detection and Localization Source: Zeiler and Fergus (2013) 18 Supervised Learning Training data has inputs (x) as well as outputs (y) Regression: When the output is real-valued, e.g.,housing data Classification: Output is a category Binary classification -- only two classes e.g.,cancer, spam Multi-class classification -- several classes e.g.,object detection 19 Unsupervised Learning : Grouping News Articles Group items into categories: sports, music, business, etc. Labels are not known Algorithm cannot know label names 20 Unsupervised Learning : Genetic Data of European Populations Source: Novembre et al., Nature (2008) 21 Active and Semi-Supervised Learning Active Learning Data is unlabelled Learning algorithm can ask for a label (from a human) Semi-supervised Learning Some data is labelled, a lot more unlabelled Can using the two together help? 22 Anomaly Detection or One-class Classification Examples Detect possible malfunction at nuclear reactors Detect fraudulent transactions for credit cards Supervised learning vs anomaly detection Anomalous events much rarer, possibly not related to each other 23 Recommendation Systems Movie / User Alice Bob Charlie Dean Eve The Shawshank Redemption The Godfather 3? The Dark Knight 5 9? 6? Pulp Fiction? 5?? 10 Schindler s List? 6? 9? Netflix competition to predict user-ratings ( ) Applications to all kinds of product recommendations No user will have used several products; take advantage of large number of users 24 Reinforcement Learning Automatic flying helicopter; self-driving cars Cannot program by hand Stochastic environment (hard to define precisely) Must take sequential decisions Can define reward functions Fun: Playing Atari breakout! [video] 25 Outline History of Machine Learning This Class Some Machine Learning Applications Some Practical Concerns Cleaning up data Spam Classification Look for words such as Nigeria, millions, Viagra, etc. Features such as the IP, other metadata If addressed by name Getting Features Often hand-crafted features by domain experts This class mainly assumes we already have features Feature learning using deep networks 26 Some pitfalls Sample To build a spam classifier, we look for words such as Nigeria, millions, etc. 27 Some pitfalls Sample To build a spam classifier, we look for words such as Nigeria, millions, etc. Training vs Test Data Future data should look like past data Not true for spam classification 27 Cats vs Dogs 28 Next Class Linear Regression Brush up your linear algebra and calculus! 29
We Need Your Support
Thank you for visiting our website and your interest in our free products and services. We are nonprofit website to share and download documents. To the running of this website, we need your help to support us.

Thanks to everyone for your continued support.

No, Thanks