1. Supervised Machine Learning: Classification Techniques Chaleece Sandberg Chris Bradley Kyle Walsh

2. Supervised Machine Learning  SML: machine performs function (e.g., classification) after training on a data set where inputs and desired outputs are provided  Following training, SML algorithm is able to generalize to new, unseen data  Application: “Data Mining” Often, large amounts of data must be handled efficiently Look for relevant information, patterns in data

3. Decision Trees  Logic-based algorithm  Sort instances (data) according to feature values…a hierarchy of tests  Nodes: features Root node: feature that best divides data  Algorithms exist for determining the best root node  Branches: values the node can assume

4. Decision Trees, an example INPUT: data (symptom) low RBC count yes no size of cells largesmall bleeding? STOP B 12 deficient? yes stop bleeding no STOPgastrin assay positive negative STOPanemia OUTPUT: category (condition)

5. Decision Trees: Assessment  Advantages:  Classification of data based on limiting features is intuitive  Handles discrete/categorical features best  Limitations:  Danger of “overfitting” the data  Not the best choice for accuracy

6. Bayesian Networks  Graphical algorithm that encodes the joint probability distribution of a data set  Captures probabilistic relationships between variables  Based on probability that instances (data) belong in each category

7. Bayesian Networks, an example Wikipedia, 2008

8. Bayesian Networks: Assessment  Advantages:  Takes into account prior information regarding relationships among features  Probabilities can be updated based on outcomes  Fast!…with respect to learning classification  Can handle incomplete sets of data  Avoids “overfitting” of data  Limitations:  Not suitable for data sets with many features  Not the best choice for accuracy

9. Neural Networks  Used for: Classification Noise reduction Prediction  Great because: Able to learn Able to generalize  Kiran Plaut’s (1996) semantic neural network that could be lesioned and retrained – useful for predicting treatment outcomes  Mikkulainen Evolving neural network that could adapt to the gaming environment – useful learning application

10. Neural Networks: Biological Basis

11. Feed-forward Neural Network Perceptron : Hidden layer

12. Neural Networks: Training  Presenting the network with sample data and modifying the weights to better approximate the desired function.  Supervised Learning Supply network with inputs and desired outputs Initially, the weights are randomly set Weights modified to reduce difference between actual and desired outputs Backpropagation

14. Support Vector Machines

15. Perceptron Revisited: Linear Classifier: y(x) = sign(w. x + b) w. x + b = 0 w. x + b < 0 w. x + b > 0

16. Which one is the best?

17. Notion of Margin  Distance from a data point to the hyperplane:  Data points closest to the boundary are called support vectors  Margin d is the distance between two classes. r d

18. Maximizing Margin  Maximizing margin is a quadratic optimization problem.  Quadratic optimization problems are a well-known class of mathematical programming problems, and many (rather intricate) algorithms exist for solving them.

19. Kernel Trick  What if the dataset is non-linearly separable?  We use a kernel to map the data to a higher-dimensional space: 0 x2x2 x 0 x

20. Non-linear SVMs: Feature spaces  General idea: The original space can always be mapped to some higher-dimensional feature space where the training set becomes separable: Φ: x → φ(x)

21. Examples of Kernel Trick  For the example in the previous figure: The non-linear mapping  A more commonly used radial basis function (RBF) kernel

23. Advantages and Applications of SVM  Advantages of SVM Unlike neural networks, the class boundaries don’t change as the weights change. Generalizability is high because margin is maximized. No local minima and robustness to outliers.  Applications of SVM Used in almost every conceivable situation where automatic classification of data is needed. (example from class) Raymond Mooney and his KRISPER natural language parser.

24. The Future of Supervised Learning (1)  Generation of synthetic data A major problem with supervised learning is the necessity of having large amounts of training data to obtain a good result. Why not create synthetic training data from real, labeled data? Example: use a 3D model to generate multiple 2D images of some object (such as a face) under different conditions (such as lighting). Labeling only needs to be done for the 3D model, not for every 2D model.

25. The Future of Supervised Learning (2)  Future applications Personal software assistants learning from past usage the evolving interests of their users in order to highlight relevant news (e.g., filtering scientific journals for articles of interest) Houses learning from experience to optimize energy costs based on the particular usage patterns of their occupants Analysis of medical records to assess which treatments are more effective for new diseases Enable robots to better interact with humans

26. References  http://homepage.psy.utexas.edu/homepa ge/class/Psy394U/Hayhoe/cognitive%20s cience%202008/talks:readings/ http://homepage.psy.utexas.edu/homepa ge/class/Psy394U/Hayhoe/cognitive%20s cience%202008/talks:readings/  http://www.ai- junkie.com/ann/evolved/nnt1.html http://www.ai- junkie.com/ann/evolved/nnt1.html  http://galaxy.agh.edu.pl/~vlsi/AI/backp_t _en/backprop.html http://galaxy.agh.edu.pl/~vlsi/AI/backp_t _en/backprop.html  http://cbcl.mit.edu/cbcl/people/heisele/h uang-blanz-heisele.pdf http://cbcl.mit.edu/cbcl/people/heisele/h uang-blanz-heisele.pdf  http://www.grappa.univ- lille3.fr/~gilleron/introML.pdf http://www.grappa.univ- lille3.fr/~gilleron/introML.pdf