Introduction Machine Learning 14022017 Machine Learning How can

Introduction Machine Learning 14/02/2017

Machine Learning How can we design a computer system whose performance improves by learning from experience?

Spam filtering

Face/person recognition demo

Recommendation systems

Robotics

Natural Language Processing

other application areas – Biometrics – Object recognition on images – DNA seqencing – Financial data mining/prediction – Process mining and optimisation Pattern Classification, Chapter 1

Big Data

Rule-based systems vs. Machine learning • Domain expert is needed for – writing rules OR – giving training sample • Which one is better? – Can the expert design rule-based systems? – Is the problem specific or general? 10

http: //www. ml-class. org/course

Most of the materials in these slides were taken from Pattern Classification (2 nd ed) by R. O. Duda, P. E. Hart and D. G. Stork, John Wiley & Sons, 2000 with the permission of the authors and the publisher

14 Definition Machine Learning (Mitchell): „a computer program 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. ”

15 Example Classify fishes see bass Classes salmon Goal: to learn a modell from training data which can categorise fishes (eg. salmons are shorter) Pattern Classification, Chapter 1

16 Classification(T) – Supervised learning: Based on training examples (E), learn a modell which works fine on previously unseen examples. – Classification: a supervised learning task of categorisation of entities into predefined set of classes Pattern Classification, Chapter 1

17 Pattern Classification, Chapter 1

Basic definitions Feature (or attribute) Instance (or entity, sample) ID Length (cm) Lightness Type 1 28 0. 5 salmon 2 23 0. 7 salmon 3 17 0. 5 sea bass Class label

19 Example Preprocessing – Image processing steps • E. g segmentation of fish contour and background – Feature extraction • Extraction of features/attributes from images which are atomic variables • Typically numerical or categorical Pattern Classification, Chapter 1

20 Example features • • • length lightness width number of paddles position of mouth Pattern Classification, Chapter 1

21 Length is a weak discriminator of fish types. Pattern Classification, Chapter 1

22 Lightness is better Pattern Classification, Chapter 1

23 Performance evaluation (P) – most simple: accuracy (correct rate) – False positive/negative errors – E. g. if the threshold is decreased the number of sea basses falsly classified to salmon decreases Decision theory Pattern Classification, Chapter 1

24 Feature vector A vector of features describing a particular instance. Instance. A x. T = [x 1, x 2] Lightness Width Pattern Classification, Chapter 1

25 Pattern Classification, Chapter 1

26 Feature space Be careful by adding to many features – noisy features (eg. measurement errors) – Unnecessary (pl. information content is similar to other feature) We need features which might have discriminative power. Feature set engineering is highly taskspecific! Pattern Classification, Chapter 1

27 This is not ideal. Remember supervised learning principle! Pattern Classification, Chapter 1

28 Pattern Classification, Chapter 1

29 Modell selection • Number of features? • Complexity of the task? • Classifier speed? • Task and data-dependent! Pattern Classification, Chapter 1

The machine learning lifecycle • • • 30 Data preparation Feature engineering Modell selection Modell training Performance evaluation Pattern Classification, Chapter 1

31 Data preparation Do we know whether we collected enough and representative sample for training a system? Pattern Classification, Chapter 1

32 Modell selection and training – These topics are the foci of this course – Investigate the data for modell selection! No free lunch! Pattern Classification, Chapter 1

33 Performance evaluation • There are various evaluation metrics • Simulation of supervised learning: 1. split your data into two parts 2. train your modell on the training set 3. predict and evaluate your modell on the test set (unknow during training) Pattern Classification, Chapter 1

34 Topics of the course • • Classification Regression Clustering Recommendation systems Learning to rank Structure prediction Reinforcement learning

https: //www. kaggle. com/competitions http: //www 195. pair. com/mik 3 hall/weka_kaggle. html