Presentation is loading. Please wait.

Presentation is loading. Please wait.

Computational Intelligence Professor J. F. Baldwin.

Similar presentations


Presentation on theme: "Computational Intelligence Professor J. F. Baldwin."— Presentation transcript:

1 Computational Intelligence Professor J. F. Baldwin

2 A Basic Data Mining Problem radiustexture perimeter area smoothness compactness concavity Concave pts symmetry Fractal dim mean Standard error largest 123456 78910 11 121314151617181920 21222324252627282930 FEATURES Breast Cancer Data Base 569 digitised images of fine needle aspirates of breast masses - 375 were benign and 212 were malignant. Use 400 training set and 169 test set Find good features and find rules to predict malignant or not

3 AI in action - 1 Flight Plan All actions and 20 variable state of aeroplane recorded during flight (simulator) of two pilots - 30 times. 90, 000 examples of (state, decision) from recordings used to obtain decision tree for action v state. Decision tree converted to program to fly plane. Program gave better performance than human Pilots.

4 AI in action - 2 PATHFINDER diagnostic expert system for lymph-node diseases. Pathfinder 1 : rule based with no account of uncertainty Pathfinder 2 : Naive Bayesian Net - gave 79% accuracy Pathfinder 3 : Bayesian Net - gave 89% accuracy Knowledge Experts designed net Performs better than world expert pathologists

5 Induction Example Ace of spades - good Queen of diamonds - good King of hearts - good Jack of clubs - good 10 of diamonds - bad 6 of clubs - bad 3 of hearts - bad 5 of spades - bad Example Set Induced Rule: card good IF card is Royal else bad Background Knowledge: Ace, King, Queen, Jack are Royal Cards We generalise from specific cases to other cases But There is no right answer

6 Induction Set of +ve and -ve Examples for a given concept Knowledge Rules for concept or other form of knowledge representation specific general Background Knowledge using background knowledge No guarantee of validity

7 More Examples 1 2 3 4 ?What is next number in sequence? etc Pictures of male faces etc Pictures of female faces male or female face What features should we use for this classification What weights of importance should we give to these features

8 General Points Features should discriminate between +ve and -ve concepts Many rules will satisfy examples Science attempts to find the simplest rules We can generalise to classification problems with more than two classes We should be able to make predictions as well as classifications

9 Vector Processing ABCD Taste Receptors on human tongue ABCDABCD Sweet Sour Salty Bitter Peach Pattern 0 10 smell Human 10 possibilities 6 0 30 Dog 30 possibilities 7

10 Message Personalisation P1P2P3P4P5 message s1s1 s2s2 s3s3 s4s4 s5s5 Personalised Fusion to user prototypes supports for interest s send ? Final Support

11 Super Market Application FreshnessProductionVitamin CFat CSalt CSuagr CFibre C Baskets Find overlapping clusters and name each cluster e.g. healthy eater, junk eater, etc. A point will have membership in each cluster Cost Each cluster represents a prototype. Use prototypes to answer such queries as - would a person be Interested in a certain product, etc.

12 Data... Knowledge Data Knowledge Decision Solution AI Tools Decision Tools Business Tools Data Mining Reporting visualisation

13 Fundamentals of AI Logic Theorem prover Graphs Search Probability Bayesian Nets Neural Nets Decision Trees Expert Systems Distributed Intelligence Problem Solving Data mining Computational Intelligence

14 An Example card 1 - bad card 2 - good card 3 - good card 4 - bad card 5 - bad card 6 - good Example set card 8 - ??? card 7 - ??? predict Are these cards Good or bad?

15 L Example Good L Bad L twice AB CD Notation Example Set unknown Predict if these unknown cases are good or bad

16 Word Blindness AB C D g1 g2 f1 f2 p {A, C}{B, D} Project onto X axis since Y axis damaged Gives word blindness since we can no longer discriminate between A and C and between B and D Humans use three features to distinguish colours. We can project in three ways onto two dimensions. This lead to three different types of colour blindness

17 Why do we need to Generalise Suppose we have a Database with 12 attributes Each attribute can take 5 possible values There are 5 12 = 2.44 10 8 possible combinations Even with a table of 1 million examples we would find an exact match only about 0.4% of the time. We use the given database to form a model which will allow us to generalise to other cases. Model can be decision tree, AI rules, Bayesian net Neural net for example

18 Fril Data Browser Relevant Data Collection Net Databases Intelligent Agents Generalisation & Summarization Inference Query & Solution Interpretation Human / Computer Interface Knowledge Base Integrity Checker Conversation Supervisor Fuzzy Inductive Logic Numeric, Text, Audio, Graphics Feature Extraction


Download ppt "Computational Intelligence Professor J. F. Baldwin."

Similar presentations


Ads by Google