CS798: Selected topics in Machine Learning Introduction Jakramate Bootkrajang Department of Computer Science Chiang Mai University Jakramate Bootkrajang CS798: Selected topics in Machine Learning 1 / 22
About the course CS789: Machine learning (and optimisation) Lecturer: Jakramate Bootkrajang Email: jakramateb@cmuacth Office hour: I am almost always at my desk, just walk in Grading: 40% homework, 30% midterm, 30% final Programming language: MATLAB, Scilab, Julia Background: Linear algebra, calculus, basic probability Jakramate Bootkrajang CS798: Selected topics in Machine Learning 2 / 22
Course outline Supervised learning methods Bayes classifier Normal discriminant analysis Logistic regression Support Vector Machine + kernel method Combining multiple classifiers: boosting, bagging Classifier evaluation Regularisation Unsupervised learning method Mixture model and EM algorithm Clustering Introduction to learning theory Learnability, PAC learning Hypothesis space, Bias, Variance Jakramate Bootkrajang CS798: Selected topics in Machine Learning 3 / 22
Human learning What is learning? In general, learning is the act of acquiring new knowledge, or modifying and reinforcing, existing knowledge and may involve synthesizing different types of information (Wikipedia) Why do we learn? Short term: Be able to do things Longer term: Money Fame Happiness How did we do? How well did we do those things How much money do we have? How happy are we? Jakramate Bootkrajang CS798: Selected topics in Machine Learning 4 / 22
Machine learning What is [machine] learning? Mathematical modelling of nature and adjusting the parameters of the model using example data or past experience (Function estimation) In some sense, it is the estimation of a function f(x) where x is the representation of each data point Why does machine learn? Short term: Be able to do some specific tasks, for example, classification or clustering Longer term: too difficult This is called learning objective How well does it perform? The machine is expected to perform well according to performance criterion For classification: right/wrong predictions For clustering: compactness of the clusters found Jakramate Bootkrajang CS798: Selected topics in Machine Learning 5 / 22
When to learn Learning is used when: Humans are unable to explain their expertise (speech recognition) Human expertise does not exist (Fraud detection) Solution changes in time (online learning, objective function changes) Otherwise, learning might not be necessary (but still possible) For example, to convert kilometre to mile Jakramate Bootkrajang CS798: Selected topics in Machine Learning 6 / 22
Steps in machine learning 1 Know what you want to do Know your data (Objective) 2 Build a model that is a good and useful approximation to the data (Modelling) 3 Devise an algorithm to learn the model: how to adjust model s parameters (Learning) 4 Test your model using existing data or new unseen data (Performance measure) 5 Theoretically show that your model will work on any new data of the same kind (Performance measure, in general) Jakramate Bootkrajang CS798: Selected topics in Machine Learning 7 / 22
Machine learning and related fields Machine learning: Focuses on modelling, learning and performance measure in general Pattern recognition: Sub-field of machine learning focuses on classification tasks Data mining: Focuses on objective and performance measure Optimisation: Focuses on learning Supporting fields: Mathematics, Statistics Jakramate Bootkrajang CS798: Selected topics in Machine Learning 8 / 22
Definition and Terminology Example : item, instance of the data used Features : attributes associated to an item, often represented as a vector (eg, word counts) Labels : category (for classification) or real value (regression) associated to an item Data : training data (labelled) test data (labelled but hidden from the learning machine) validation data (labelled, for tuning parameters) Jakramate Bootkrajang CS798: Selected topics in Machine Learning 9 / 22
Paradigms of machine learning Supervised learning: the task of inferring a function from labelled data (classification and regression) Unsupervised learning: the task of inferring a function from unlabelled data to describe hidden structure (clustering) Semi-supervised learning: inferring a function from labelled and unlabelled data Reinforcement learning: the task of inferring a function from interaction with the world based on awards and penalty Correct input/output pairs are never presented (Robot control) Jakramate Bootkrajang CS798: Selected topics in Machine Learning 10 / 22
Classification Credit scoring Differentiating beween low-risk and high-risk customers from their income and savings Discriminant IF income > θ 1 AND savings > θ 2 THEN low-risk ELSE high risk Jakramate Bootkrajang CS798: Selected topics in Machine Learning 11 / 22
Classification: applications Also known as: pattern recognition Face recognition: Pose, lighting, occlusion (glasses, beard), make-up, hair style Character recognition: Different handwriting styles Speech recognition: Temporal dependency Medical diagnosis: From symptoms to illnesses Jakramate Bootkrajang CS798: Selected topics in Machine Learning 12 / 22
Face recognition Jakramate Bootkrajang CS798: Selected topics in Machine Learning 13 / 22
Regression Price of a used car y = g(x θ), x: car attributes, y: price g() is the model and θ is model s parameter For navigating a car: Angle of the steering wheel as output (CMU NavLab) Jakramate Bootkrajang CS798: Selected topics in Machine Learning 14 / 22
Unsupervised learning Learning what normally happens Clustering: Grouping similar instances Example applications Customer segmentation in CRM Image compression: Color quantization Jakramate Bootkrajang CS798: Selected topics in Machine Learning 15 / 22
Semi-supervised learning Learning from small set of labelled data + large amount of unlabelled data Focus on how to make use of unlabelled data to improve the performance Possible applications are very similar to supervised learning Example: Lanna OCR Jakramate Bootkrajang CS798: Selected topics in Machine Learning 16 / 22
Reinforcement learning Learning a policy: A sequence of outputs (actions) Reward good behaviour Punish bad behaviour Many control related applications Game playing Robot in a maze Some demo Jakramate Bootkrajang CS798: Selected topics in Machine Learning 17 / 22
Resources: Books Foundation of machine learning, Mehryar Mohri, Afshin Rostamizadeh and Ameet Talwalkar, The MIT Press Pattern classification, Richard Duda, Peter Hart, David Storck Wiley-Interscience Understanding machine learning, Shai Shalev-Shwartz, Shai Ben-David, Cambridge University Press Machine learning, Tom Mitchell, McGraw-Hill Education Pattern recognition and machine learning, Christopher Bishop, Springer Jakramate Bootkrajang CS798: Selected topics in Machine Learning 18 / 22
Resources: Datasets UCI Repository: http://archiveicsuciedu/ml/ UCI KDD Archive: https://kddicsuciedu/ Delve: http://wwwcstorontoedu/ delve/data/datasetshtml Pascal Large Scale Learning Challenge: http://largescalemltu-berlinde/about/ ImageNet: http://image-netorg/index Jakramate Bootkrajang CS798: Selected topics in Machine Learning 19 / 22
Prominant figures Michael I Jordan (UC Berkeley) (Bayesian inference) Bernhard Schölkopf (TU Berlin) (Kernel methods) Chris Williams (University of Edinburgh) (Gaussian process) John Shawe-Taylor (UCL) (Theory) Andrew Ng (Stanford) Geoffrey Hinton (University of Toronto) (Deep learning) Carl Edward Rasmussen, Christopher M Bishop, Zoubin Ghahramani, Terrence Sejnowski Jakramate Bootkrajang CS798: Selected topics in Machine Learning 20 / 22
Journals Journal of Machine Learning Research (JMLR) Machine Learning Neural Computation Neural Networks IEEE Transactions on Neural Networks IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI) Pattern Recognition Neurocomputing Jakramate Bootkrajang CS798: Selected topics in Machine Learning 21 / 22
Conferences International Conference on Machine Learning (ICML) European Conference on Machine Learning (ECML) Neural Information Processing Systems (NIPS) Uncertainty in Artificial Intelligence (UAI) Computational Learning Theory (COLT) International Joint Conference on Artificial Intelligence (IJCAI) International Conference on Neural Networks (Europe) Jakramate Bootkrajang CS798: Selected topics in Machine Learning 22 / 22