![What is all the hype about ML? Why machine learning? Typical Data Mining Task
data mining: pattern recognition, knowledge
discovery, use historical data to improve future
"Every time I fire a linguist the performance of the decisions, prediction (classification, regression),
recognizer goes up" data discripton (clustering, summarization,
visualization)
(probably) said by Fred Jelinek (IBM speech group) in the 80s, quoted by, e.g., complex applications: we cannot program by hand,
Jurafsky and Martin, Speech and Language Processing. (efficient) processing of complex signals Given:
self-customizing programs: automatic adjustments • 9714 patient records, each describing a pregnancy and birth
according to usage, dynamic systems • Each patient record contains 215 features
Learn to predict:
• Classes of future patients at high risk for Emergency Cesarean Section
Machine Learning, LIX004M5 – p.10/50 Machine Learning, LIX004M5 – p.11/50 Machine Learning, LIX004M5 – p.12/50
Pattern Recognition Complex applications Classification
Object detection Operating robots: Personal home page? Company website? Educational site?
ALVINN [Pomerleau] drives 70 mph on highways
Machine Learning, LIX004M5 – p.13/50 Machine Learning, LIX004M5 – p.14/50 Machine Learning, LIX004M5 – p.15/50
Automatic customization Machine learning is growing Questions to ask
many more applications: Learning = improve with experience at some task
• speech recognition • What experience?
• robot control • What exactly should be learned?
• spam filtering, data sorting • How shall it be represented?
• machine translation • What specific algorithm to learn it?
• financial data analysis and market predictions Goal: handle unseen data correctly according to the
• hand writing recognition task (use your knowledge inferred from experience!)
• data clustering and visualization
• pattern recognition in genetics (e.g. DNA
sequences)
Machine Learning, LIX004M5 – p.16/50 Machine Learning, LIX004M5 – p.17/50 Machine Learning, LIX004M5 – p.18/50](https://image.slidesharecdn.com/machine-learning-lix004m52561/85/Machine-Learning-LIX004M5-2-320.jpg)
Machine Learning, LIX004M5
- 1. General Info General Info (cont’d) instructor: Jörg Tiedemann (j.tiedemann@rug.nl) Machine Learning, LIX004M5 Harmoniegebouw, room 1311-429 You need an account on ’hagen’ for the labs! (But you may also work from home or somewhere Overview and Introduction prerequisites: open to students in Computer else) Science, Artificial Intelligence and Information J¨ rg Tiedemann o Science tiedeman@let.rug.nl 2nd year student or higher background: programming ability, elementary Go to A. Da Costa. (Harmoniegebouw, k. 336, Informatiekunde Rijksuniversiteit Groningen statistics bouwdeel 13.13, telefoon 363 5801) (dagelijks 10.30- schedule: September 5 - October 21 12.00 en 14.00-15.30 uur, vrijdagmiddag gesloten)! • lectures: mondays 13-15 • labs: fridays 9-11 (4 times only!) Machine Learning, LIX004M5 – p.1/50 Machine Learning, LIX004M5 – p.2/50 Machine Learning, LIX004M5 – p.3/50 General Info (cont’d) General Info (cont’d) General Info (cont’d) Website: http://www.let.rug.nl/˜tiedeman/ml06 Purpose of this course • Examination: Examination: lab assignments and exercises (50%), • Introduction (!) to machine learning techniques • obligatory lab assignments (50%) written exam (50%) (How much do you know already?) • written exam (50%) Exam: Friday, October 27, 9-12 • Discussion of several machine learning • A minimum of 6 for both parts is required! Literature: Tom Mitchell Machine Learning, New approaches • Exam is open book York: McGraw-Hill, 1997 • Examples and applications in various fields additional on-line literature (links available from • Practical assignments the course website) • using Weka - a machine learning package implemented in Java • a little bit of programming/scripting • some theoretical questions Machine Learning, LIX004M5 – p.4/50 Machine Learning, LIX004M5 – p.5/50 Machine Learning, LIX004M5 – p.6/50 Preliminary Program General comments What is Machine learning? 1. Organization, Introduction, (Ch.1, Ch.5) • Read the book! (and other literature if necessary) Machine Learning is 2. Inductive learning (Ch.2), Decision Trees (Ch.3) • Ask questions! (and I’ll try to answer) • the study of algorithms that • Lab 1 - Decision Trees • Tell me if you think that something’s wrong • improve their performance 3. Instance-Based Learning (Ch.8) • at some task • Keep the deadlines! • Lab 2 - Instance-based learning (1 week late → half the points, later → no points) • with experience 4. Bayesian Learning (Ch.6) • Lab 3 - Learner comparison/combination ... just like a human being ... (?) 5. Sequential data & Markov Models, M&S Ch.9, Bilmex; • Lab 4 - Markov models 6. Maximum Entropy models, Combining Learners 7. Genetic Algorithms (Ch.9), Reinforcement Learning (Ch.13) Machine Learning, LIX004M5 – p.7/50 Machine Learning, LIX004M5 – p.8/50 Machine Learning, LIX004M5 – p.9/50
- 2. What is all the hype about ML? Why machine learning? Typical Data Mining Task data mining: pattern recognition, knowledge discovery, use historical data to improve future "Every time I fire a linguist the performance of the decisions, prediction (classification, regression), recognizer goes up" data discripton (clustering, summarization, visualization) (probably) said by Fred Jelinek (IBM speech group) in the 80s, quoted by, e.g., complex applications: we cannot program by hand, Jurafsky and Martin, Speech and Language Processing. (efficient) processing of complex signals Given: self-customizing programs: automatic adjustments • 9714 patient records, each describing a pregnancy and birth according to usage, dynamic systems • Each patient record contains 215 features Learn to predict: • Classes of future patients at high risk for Emergency Cesarean Section Machine Learning, LIX004M5 – p.10/50 Machine Learning, LIX004M5 – p.11/50 Machine Learning, LIX004M5 – p.12/50 Pattern Recognition Complex applications Classification Object detection Operating robots: Personal home page? Company website? Educational site? ALVINN [Pomerleau] drives 70 mph on highways Machine Learning, LIX004M5 – p.13/50 Machine Learning, LIX004M5 – p.14/50 Machine Learning, LIX004M5 – p.15/50 Automatic customization Machine learning is growing Questions to ask many more applications: Learning = improve with experience at some task • speech recognition • What experience? • robot control • What exactly should be learned? • spam filtering, data sorting • How shall it be represented? • machine translation • What specific algorithm to learn it? • financial data analysis and market predictions Goal: handle unseen data correctly according to the • hand writing recognition task (use your knowledge inferred from experience!) • data clustering and visualization • pattern recognition in genetics (e.g. DNA sequences) Machine Learning, LIX004M5 – p.16/50 Machine Learning, LIX004M5 – p.17/50 Machine Learning, LIX004M5 – p.18/50
- 3. What experience? What exactly should be learned? How shall it be represented? • What do we know already about the task and Outcome of the target function Model selection possible solutions? (prior knowledge) • boolean (→ concept learning) • symbolic representation (e.g. rules) • What kind of data do we have available? • discrete values (→ classification) (training examples) • subsymbolic representation (neural networks, • real values (→ regression) SVMs) What are the discriminative features? How are they connected with each other (dependencies)? many machine learning tasks are classification tasks ... Do we want to restrict the space of possible solutions? • Is a “teacher” available (→ supervised learning) (→ restriction bias ... we come back to this) or not (→ unsupervised learning)? How expensive is labeling? • How much data do we need and how clean does it have to be? Machine Learning, LIX004M5 – p.19/50 Machine Learning, LIX004M5 – p.20/50 Machine Learning, LIX004M5 – p.21/50 What algorithm to learn it? What algorithm to learn it? Learning Models Learning means approximating the real (unknown) Learning means approximating the real (unknown) Learning means approximating the real (unknown) target function according to our experience (e.g. target function according to our experience (e.g. target function according to our experience (e.g. observed training examples) observed training examples) observed training examples) → Learning = Search for a “good” hypothesis/model → Learning = search for a “good” hypothesis/model Do we want to prefer certain models? Which one is better? (→ preference bias ... later more) Machine Learning, LIX004M5 – p.22/50 Machine Learning, LIX004M5 – p.23/50 Machine Learning, LIX004M5 – p.24/50 Learning Models What algorithm to learn it? The roots of ML Learning means approximating the real (unknown) Learning means approximating the real (unknown) Artificial intelligence: use prior knowledge and training data to guide target function according to our experience (e.g. target function according to our experience (e.g. learning as a search problem observed training examples) observed training examples) Baysian methods: probabilistic classifiers, probabilistic reasoning Computational complexity theory: trade-off between model (learning) → Learning = search for a “good” hypothesis/model → Learning = search for a “good” hypothesis/model complexity and performance Control theory: control optimisation processes Which one is better? • supervised learning (classified data available) • unsupervised learning (e.g. clustering) Information theory: entropy, information content, code optimsation and the minimum description length principle • inductive learning (from training data) Philosophy: Occam’s razor (simple is best) • deductive learning (data + domain theory) Psychology and neurobiology: response improvement with practice, ideas • gradient descent, bayesian learning, reinforcement learning ... that lead to artificical neural networks Statistics: data description, estimation of probability distributions, evaluation, confidence Machine Learning, LIX004M5 – p.25/50 Machine Learning, LIX004M5 – p.26/50 Machine Learning, LIX004M5 – p.27/50
- 4. A walk-through example A walk-through example A walk-through example from Duda et al: Pattern Classification from Duda et al: Pattern Classification Procedure: • Task: automatically sort incoming fish on a • Task: automatically sort incoming fish on a preprocessing: isolate fishes from one another and conveyor belt into “sea bass” or “salmon” conveyor belt into “sea bass” or “salmon” from the background of the images • Experience: sample images feature selection: determine discriminative features We want a machine to learn this task. The machine needs some “experience”. to be extracted from the images (e.g. length, lightness, width, position of mouth, etc), feature selection = kind of data reduction (focus on relevant information) feature extraction: extract selected features from images and pass them to a classifier Machine Learning, LIX004M5 – p.28/50 Machine Learning, LIX004M5 – p.29/50 Machine Learning, LIX004M5 – p.30/50 A walk-through example A walk-through example A walk-through example Select length for discrimiation: Lightness is a better feature: Machine Learning, LIX004M5 – p.31/50 Machine Learning, LIX004M5 – p.32/50 Machine Learning, LIX004M5 – p.33/50 A walk-through example A walk-through example A walk-through example • devise decision rule or move decision boundary Feature selection: We still need to: to minimize some classification cost (→ decision • distinguishing (similar for objects in same theory) • select an appropriate type of model for category and very different for different classification (e.g. function class to define • a single feature might not be enough minimize categories) separation boundaries) costs • invariant (feature value doesn’t change when • select the model that generalizes the best (to be → feature vector, e.g.: changing the context) able to classify even unseen objects correctly) • insensitive to noise • consider computational complexity (trade-off X1 width • simple to extract between complexity and performance; scalibility) X= X2 lightness Machine Learning, LIX004M5 – p.34/50 Machine Learning, LIX004M5 – p.35/50 Machine Learning, LIX004M5 – p.36/50
- 5. A walk-through example A walk-through example A walk-through example Linear decision boundary Overly complex decision boundary a good trade-off between performance on the training (What is the problem?) set and model simplicity Machine Learning, LIX004M5 – p.37/50 Machine Learning, LIX004M5 – p.38/50 Machine Learning, LIX004M5 – p.39/50 The Design Cycle Evaluation Evaluation We have ... Evaluation of classifiers based on • different feature sets • accuracy or error rate • different models (percentage of classification errors) • different learning strategies • risk (cost estimation for classification decisions) → We need to evaluate! Never ever evaluate on training data! ... Why not? Machine Learning, LIX004M5 – p.40/50 Machine Learning, LIX004M5 – p.41/50 Machine Learning, LIX004M5 – p.42/50 Evaluation Evaluation Evaluation Typical strategy in supervised learning: Distinguish: How good is an estimate of the true error by means of Split data into disjoint training data and test data sample errors? sample error: error rate observed when classifying Problems: sample data (test data) • confidence intervals true error: probability of misclassifying a randomly • larger sample → greater confidence • we could be (too) lucky (sample error on test data selected object is better than with other data splits) • test data set is too small to be confident How good is one model compared to another? • training data is rare and expensive (we don’t want • calculate sample errors to waste too much when seperating test data) • compute statistical significance (e.g. paired t-test) Machine Learning, LIX004M5 – p.43/50 Machine Learning, LIX004M5 – p.44/50 Machine Learning, LIX004M5 – p.45/50
- 6. Cross validation Cross validation Cross validation • split D into k similar sized sets (e.g. k=10) • split D into k similar sized sets (e.g. k=10) • split D into k similar sized sets (e.g. k=10) • use k − 1 sets for training and 1 for evaluation • use k − 1 sets for training and 1 for evaluation • use k − 1 sets for training and 1 for evaluation • use each set once for evaluation and calculate the • use each set once for evaluation and calculate the • use each set once for evaluation and calculate the average of the errors average of the errors average of the errors → improve error estimates (higher confidence) → improve error estimates (higher confidence) → all data is tested → all data is tested → better use of (limited) training data → better use of (limited) training data Note: we still don’t evaluate on training data! Note: we still don’t evaluate on training data! special case: leave one out cross validation - use each training example once for testing and the others for training Machine Learning, LIX004M5 – p.46/50 Machine Learning, LIX004M5 – p.47/50 Machine Learning, LIX004M5 – p.48/50 Conclusion What’s next? • we seem to be overwhelmed by number, complexity and magnitude of sub-problems This week: Read ch. 1 & ch. 5 of Mitchell and look • many of them can be solved (to a certain degree at the exercises at least) No lab on Friday! • many fascinating problems still remain Next week: Inductive learning, Mitchell ch. 2 & Enjoy working with learning systems! Decision trees, ch. 3 First lab about Decision Trees Machine Learning, LIX004M5 – p.49/50 Machine Learning, LIX004M5 – p.50/50