Chapter 11 Analysis and Explanation. Chapter 11 Outline Explain how CI systems do what they do Only a few methodologies are discussed here Sensitivity.

Slides:



Advertisements
Similar presentations
Multi-Layer Perceptron (MLP)
Advertisements

Design of Experiments Lecture I
NEURAL NETWORKS Backpropagation Algorithm
1 Machine Learning: Lecture 4 Artificial Neural Networks (Based on Chapter 4 of Mitchell T.., Machine Learning, 1997)
CHAPTER 13 Inference Techniques. Reasoning in Artificial Intelligence n Knowledge must be processed (reasoned with) n Computer program accesses knowledge.
Particle Swarm Optimization (PSO)  Kennedy, J., Eberhart, R. C. (1995). Particle swarm optimization. Proc. IEEE International Conference.
Rule Based Systems Michael J. Watts
Artificial Neural Networks
Machine Learning Neural Networks
Simple Neural Nets For Pattern Classification
Supervised learning 1.Early learning algorithms 2.First order gradient methods 3.Second order gradient methods.
The back-propagation training algorithm
Prénom Nom Document Analysis: Artificial Neural Networks Prof. Rolf Ingold, University of Fribourg Master course, spring semester 2008.
Connectionist models. Connectionist Models Motivated by Brain rather than Mind –A large number of very simple processing elements –A large number of weighted.
September 30, 2010Neural Networks Lecture 8: Backpropagation Learning 1 Sigmoidal Neurons In backpropagation networks, we typically choose  = 1 and 
Neural Networks Marco Loog.
Neural Networks. R & G Chapter Feed-Forward Neural Networks otherwise known as The Multi-layer Perceptron or The Back-Propagation Neural Network.
Prénom Nom Document Analysis: Artificial Neural Networks Prof. Rolf Ingold, University of Fribourg Master course, spring semester 2008.
Neural Networks Chapter Feed-Forward Neural Networks.
1 Validation and Verification of Simulation Models.
Classification and Prediction by Yen-Hsien Lee Department of Information Management College of Management National Sun Yat-Sen University March 4, 2003.
October 14, 2010Neural Networks Lecture 12: Backpropagation Examples 1 Example I: Predicting the Weather We decide (or experimentally determine) to use.
CHAPTER 11 Back-Propagation Ming-Feng Yeh.
CS 4700: Foundations of Artificial Intelligence
Hub Queue Size Analyzer Implementing Neural Networks in practice.
Neural Networks. Background - Neural Networks can be : Biological - Biological models Artificial - Artificial models - Desire to produce artificial systems.
Ranga Rodrigo April 5, 2014 Most of the sides are from the Matlab tutorial. 1.
Dr. Hala Moushir Ebied Faculty of Computers & Information Sciences
Artificial Neural Networks (ANN). Output Y is 1 if at least two of the three inputs are equal to 1.
Artificial Neural Networks
CSC2535: Computation in Neural Networks Lecture 11: Conditional Random Fields Geoffrey Hinton.
Chapter 9 Neural Network.
Chapter 3 Neural Network Xiu-jun GONG (Ph. D) School of Computer Science and Technology, Tianjin University
Appendix B: An Example of Back-propagation algorithm
Chapter 9 – Classification and Regression Trees
CONTENTS:  Introduction  What is neural network?  Models of neural networks  Applications  Phases in the neural network  Perceptron  Model of fire.
NEURAL NETWORKS FOR DATA MINING
 Diagram of a Neuron  The Simple Perceptron  Multilayer Neural Network  What is Hidden Layer?  Why do we Need a Hidden Layer?  How do Multilayer.
For games. 1. Control  Controllers for robotic applications.  Robot’s sensory system provides inputs and output sends the responses to the robot’s motor.
Artificial Intelligence Techniques Multilayer Perceptrons.
1 Pattern Recognition: Statistical and Neural Lonnie C. Ludeman Lecture 21 Oct 28, 2005 Nanjing University of Science & Technology.
VI. Evaluate Model Fit Basic questions that modelers must address are: How well does the model fit the data? Do changes to a model, such as reparameterization,
Chapter 1 Introduction to Statistics. Statistical Methods Were developed to serve a purpose Were developed to serve a purpose The purpose for each statistical.
Artificial Intelligence Chapter 3 Neural Networks Artificial Intelligence Chapter 3 Neural Networks Biointelligence Lab School of Computer Sci. & Eng.
Soft Computing Lecture 19 Part 2 Hybrid Intelligent Systems.
Fuzzy Systems Michael J. Watts
CSE & CSE6002E - Soft Computing Winter Semester, 2011 Neural Networks Videos Brief Review The Next Generation Neural Networks - Geoff Hinton.
Introduction to Neural Networks. Biological neural activity –Each neuron has a body, an axon, and many dendrites Can be in one of the two states: firing.
Neural Network Basics Anns are analytical systems that address problems whose solutions have not been explicitly formulated Structure in which multiple.
Introduction to Linear Algebra Mark Goldman Emily Mackevicius.
1 Statistics & R, TiP, 2011/12 Neural Networks  Technique for discrimination & regression problems  More mathematical theoretical foundation  Works.
Chapter 8: Adaptive Networks
Neural Networks Vladimir Pleskonjić 3188/ /20 Vladimir Pleskonjić General Feedforward neural networks Inputs are numeric features Outputs are in.
CHEE825 Fall 2005J. McLellan1 Nonlinear Empirical Models.
Neural Networks The Elements of Statistical Learning, Chapter 12 Presented by Nick Rizzolo.
 Problem Analysis  Coding  Debugging  Testing.
Data Mining: Concepts and Techniques1 Prediction Prediction vs. classification Classification predicts categorical class label Prediction predicts continuous-valued.
A Presentation on Adaptive Neuro-Fuzzy Inference System using Particle Swarm Optimization and it’s Application By Sumanta Kundu (En.R.No.
Evolutionary Computation Evolving Neural Network Topologies.
Chapter 12 Case Studies Part B. Control System Design.
The Gradient Descent Algorithm
第 3 章 神经网络.
Neural Networks: Improving Performance in X-ray Lithography Applications ECE 539 Ryan T. Hogg May 10, 2000.
Dr. Unnikrishnan P.C. Professor, EEE
of the Artificial Neural Networks.
Artificial Intelligence Chapter 3 Neural Networks
Capabilities of Threshold Neurons
Artificial Intelligence Chapter 3 Neural Networks
Artificial Intelligence Chapter 3 Neural Networks
Artificial Intelligence Chapter 3 Neural Networks
Presentation transcript:

Chapter 11 Analysis and Explanation

Chapter 11 Outline Explain how CI systems do what they do Only a few methodologies are discussed here Sensitivity analysis –Relation factors –Zurada sensitivity analysis Hinton diagrams Applications of EC to explanation facilities

Sensitivity Analysis Provides a method for assessing relative importance of CI system inputs One definition is true positive ratio: TP/(TP+FN) Another approach is that of relation factors, which depict strengths between individual inputs and individual outputs of a CI system Still another approach is the Zurada sensitivity analysis, designed originally for neural networks

Relation Factor One Relation factor one: Effect of a given input on a given output when other inputs are held constant (often 0); switch input over dynamic range (0–1 for NN, dynamic range for fuzzy system) With i inputs and o outputs, there are i * o relation factors one Can clamp values to something other than 0, say 0.5 (or the midpoint of the dynamic range, for fuzzy systems).

Relation Factor Two Measures average effect of given input on given output over a set of patterns For each pattern, calculate the change in the output when the input is switched over its range while other inputs have value defined by the pattern Repeat for all patterns Sum of changes divided by number of patterns gives the factor for a given input-output pair Again, there are i * o such factors Can use relation factors to make CI system more intelligent about what input is requested next

Zurada Sensitivity Analysis The sensitivity of a trained output z kj with respect to an input a ki is defined as The sensitivity must therefore be determined for each input for each pattern, resulting in a sensitivity matrix.

Zurada Sensitivity Three sensitivity measures are defined over the entire training set: The mean square average sensitivity matrix S avg is defined as: The absolute value average sensitivity matrix S abs is defined as: The maximum sensitivity matrix S max is defined as:

Zurada Sensitivity, Simplified Calculate mean value for each input parameter Hold all but 1 input at mean, vary other over dynamic range in steps (10-15 steps), then iterate for each input. Sensitivity of an input with respect to an output is max – min over this range of inputs: S ji,e Calculate a sensitivity for each input S i,e in one of three ways.

Zurada Sensitivity, Simplified Now calculate a sensitivity for each input The mean square average estimated sensitivity S i,eav is defined as: The absolute value average estimated sensitivity S i,eab is defined as: The maximum estimated sensitivity S i,emx is defined as: in one of three ways:

Using Zurada Sensitivities Rank order sensitivities Delete input with lowest sensitivity and retrain network If results are good, keep result and try deleting another input More scientific approach: –Retrain network with a random variable as an additional input –Calculate Zurada sensitivities –Remove any input with sensitivity lower that that of random input –Retrain (without random input) Same method can be adapted for fuzzy systems

Hinton Diagrams Must exercise care when interpreting weights in a neural network (large weights aren’t always important) Numeric representations of weight matrixes are difficult to interpret Geoffrey Hinton developed a graphical representation technique Size of shape is magnitude of weight Color or shading represents sign In Figure, a backprop net weight matrix is shown Input to hidden weights on top Hidden to output weights on bottom Bias weights on left A number of variations exist (activation values can be displayed, for example; can be used to prune networks

Hinton diagram for a feedforward neural network

EC Tools for Explanation Facilities Explanation facilities make CI systems understandable to users Explanation facilities should have consistent user interfaces Functions can include: Cite reasons for decision Make system actions clear Provide examples Cite logical relationships

Explanation Facility Justification Main justification often is to provide reasons for system conclusions Also sometimes justified by need for info on: System limitations System knowledge domain(s) Codebook vectors Decision hypersurface information The bottom line is that the user wants to TRUST the system!

Explanation Facility Design and Functions Design of interface important Design should be responsive to level of users, especially novices Trace functions used mainly for debugging NN explanation facilities can provide user with “codebook vectors” which are quintessential examples (online or offline) NN facilites can also provide information on decision hypersurface including distance to it Fuzzy System facilities can list rules that fired, ranked by contribution

Explanation Facility Shortcomings Sequence of rule firings not intuitive for many users Typical backward chaining system doesn’t give information on decision hypersurface Some explanation facilities require rule firing information Systems that have parallel aspects, such as evolutionary fuzzy expert systems, present special challenges

Evolutionary Computation Tools Use trained NN weight matrix to calculate fitness and EA to find input patterns that illustrate: Codebook vectors Decision hypersurface Some kind of rank ordering of EA is often beneficial Fuzzy systems can also act as fitness functions

Modular Approach to Explanation Facilities “Look and feel” should be consistent among modules despite using codebook vectors, relation factors, etc. A (fuzzy) rule-based shell can provide a common interface and consistency

Modular Medical Diagnostic System Could represent three main modules: abdominal disorders, chest pain, and ocular complaints.

Example Neural Network Explanation Facility Uses particle swarm optimization Works on the Iris data set Can be used with any back-propagation neural net weight file obtained using the back-propagation implementation in this book Run it by invoking the program with two run files: nnexp bp.run pso.run

PSO Run File Similar to that for the evolutionary NN application Example: = minimize 18 Use evaluation function

BP.RUN Example iris.wts neural network weight file 3 number of layers in NN 4 number of PE in hidden layer 4 number of network inputs 3 number of network outputs what you are looking for examples of acceptable sum-squared error irisexp.out results file for output Note: Don’t use 1 and 0 as your targets with a sigmoidal activation function.

Sample Output Format: inp1 inp2 inp3 inp4 targval1 targval2 targval3 error To get values near the decision hypersurface for classes 2 and 3, use target values of: