symbolic rules extraction from trained neural networks
DESCRIPTION
TRANSCRIPT
![Page 1: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/1.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 11
Symbolic Rules Symbolic Rules Extraction From Extraction From Trained Neural Trained Neural
Networks Networks
Koushal Kumar
M .Tech CSE
Mob: +918968939621
![Page 2: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/2.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 22
What are Artificial Neural Networks?What are Artificial Neural Networks?
• Artificial Neural Networks are powerful computational systems consisting of many simple processing elements connected together to perform tasks analogously to biological brains.
• They are massively parallel, which makes them efficient, robust, fault tolerant and noise tolerant
• They can learn from training data and generalize to new situations..
![Page 3: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/3.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 33
The limitation of Neural NetworkThe limitation of Neural Network
The major criticism against Neural Network is that The major criticism against Neural Network is that
decision given by neural networks is difficult to decision given by neural networks is difficult to
understand by a human being. Reasons for this areunderstand by a human being. Reasons for this are
Knowledge in Neural Networks are stored as real Knowledge in Neural Networks are stored as real
valuesvalues
parameters (weights and bias) of networksparameters (weights and bias) of networks
Neural Networks are unable to explain its internal Neural Networks are unable to explain its internal
processing how they come to particular decisionprocessing how they come to particular decision
This behavior makes Neural Networks Black Box in This behavior makes Neural Networks Black Box in
NatureNature
![Page 4: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/4.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 44
Continue..Continue..
![Page 5: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/5.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 55
Rules extraction From Rules extraction From Neural Networks.Neural Networks.
So to overcome the Black Box nature of Neural So to overcome the Black Box nature of Neural
Networks we need to extract rules from Neural Networks we need to extract rules from Neural
Networks so that the user can gain a better Networks so that the user can gain a better
understanding of the decision process. understanding of the decision process.
following types of rules can be extracted from following types of rules can be extracted from
neural networksneural networks
I) M OF N types rulesI) M OF N types rules
II) Fuzzy rulesII) Fuzzy rules
![Page 6: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/6.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 66
continue..continue..
III) III) IF THEN RULESIF THEN RULES
IV) Decision RulesIV) Decision Rules
V) First order logic rulesV) First order logic rules
From all above types of rules IF THEN RULES From all above types of rules IF THEN RULES
and Decision rules are easy to understand and Decision rules are easy to understand
then others kind of rules.then others kind of rules.
![Page 7: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/7.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 77
J48 Algorithm for extracting J48 Algorithm for extracting decision treesdecision trees
• J48 is an algorithm used to generate a decision tree.J48 is an algorithm used to generate a decision tree.• Developed by quinlan and most widely used decision Developed by quinlan and most widely used decision
tree induction algorithm.tree induction algorithm.• It is based upon greedy search approach i.e select the It is based upon greedy search approach i.e select the
best attribute and never looks back to reconsider early best attribute and never looks back to reconsider early choices.choices.
• It select the best attribute according to its entropy It select the best attribute according to its entropy value.value.
• More preference will be given to that attribute which More preference will be given to that attribute which has more value of entropy.has more value of entropy.
![Page 8: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/8.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 88
Before normalizationBefore normalization
![Page 9: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/9.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 99
After normalizationAfter normalization
![Page 10: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/10.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1010
MATLAB SimulatorMATLAB Simulator
• Matlab stands for matrix laboratory.• It integrate computation, visualization, and
programming in an easy-to-use environment.• MATLAB is a package that has been purpose-
designed to make computations easy, fast and reliable.
• Matlab can be used in math and computation, algorithm development, simulation purposes.
• MATLAB is a powerful system that can plot graphs and perform a large variety of calculations with numbers.
![Page 11: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/11.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1111
Weka simulatorWeka simulator
• WEKA is abbreviation of Waikato Environment for Knowledge Analysis.
• Weka is open source simulator with machine learning algorithms.
• The Weka workbench contains a collection of visualization tools and algorithms for data analysis and predictive modelling..
![Page 12: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/12.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1212
Training data set and its HistogramTraining data set and its Histogram
![Page 13: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/13.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1313
Histogram of input target valuesHistogram of input target values
![Page 14: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/14.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1414
Histogram of testing set as InputsHistogram of testing set as Inputs
![Page 15: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/15.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1515
Histogram of testing set as Target valuesHistogram of testing set as Target values
![Page 16: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/16.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1616
Neural networks Training with BPANeural networks Training with BPA
![Page 17: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/17.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1717
Best Training Performance curveBest Training Performance curve
![Page 18: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/18.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1818
Gradient Performance Curve After Training & Gradient Performance Curve After Training & TestingTesting
![Page 19: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/19.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 1919
Data import window in wekaData import window in weka
![Page 20: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/20.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2020
J48 algorithm Output (pruned)J48 algorithm Output (pruned)
![Page 21: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/21.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2121
Results from pruned TreeResults from pruned Tree
![Page 22: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/22.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2222
J48 algorithm Output (Unpruned)J48 algorithm Output (Unpruned)
![Page 23: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/23.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2323
Unpruned Tree ResultsUnpruned Tree Results
![Page 24: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/24.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2424
Results of classification from Decision TreeResults of classification from Decision Tree
![Page 25: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/25.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2525
Decision Tree Extracted From Exp no 7Decision Tree Extracted From Exp no 7
![Page 26: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/26.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2626
The Following Rule Set Is Obtained From the above Decision The Following Rule Set Is Obtained From the above Decision
TreeTree I. Applying Remove Redundancy ConditionsI. Applying Remove Redundancy Conditions IF Children ≥ 1 AND Children >2AND Children >3 THEN Marital status =YES
Children ≥ 1 is more specific than Children >3 and Children > 2. So we remove all such conditions
Rule 1:
a) IF Current_act = NO AND Age ≤ 48.0 AND Sex = FEMALE AND Children ≤ 0 THEN Region Town
b) IF AGE > 48.0 AND Region Suburban AND Current_act = NO then Pep = NO
II. For every pair decision trees Remove redundancy rules. For example
Rule 1: IF Age ≤60 AND Salary ≤ 3500 AND Pep = NO THEN Mortage = YES
Rule 2: IF Age ≤ 50 AND Salary ≤ 3500 AND Pep = NO THEN Mortage = YES
New Rule: IF Age ≤ 50 AND Salary ≤ 3500 AND Pep = NO THEN Mortage =YES
III. Remove more specific rules. The rules with a condition set which is a superset of
another rule should be removed. For example
![Page 27: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/27.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2727
Continue….Continue….
• Rule 1: IF Age ≤ 60 AND Region = Rural AND Saving_ act = YES THEN Pep = NO
• Rule 2: IF Age <= 60 AND Children <= 1 AND Region =Rural AND saving act = YES THEN Pep= NO
• Rule 3: IF Region = Rural AND saving _ act =YES THEN Pep = NO
• New Rule: IF Region = Rural AND saving _ act =YES THEN Pep = NO
![Page 28: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/28.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2828
Comparison of J48 algorithm with Others ClassifiersComparison of J48 algorithm with Others Classifiers
![Page 29: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/29.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 2929
Graphical comparisons of algorithmsGraphical comparisons of algorithms
![Page 30: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/30.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 3030
Data Flow diagram of J48 algorithmData Flow diagram of J48 algorithm
![Page 31: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/31.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 3131
Paper PublishedPaper Published
• ““Extracting Explanation From Artificial Neural Networks” is Extracting Explanation From Artificial Neural Networks” is published in International Journal of Computer Science and published in International Journal of Computer Science and Information Technologies.Information Technologies.
• ““Advanced Applications of Neural Networks and Artificial Advanced Applications of Neural Networks and Artificial Intelligence: A Review” has been selected in International journal of Intelligence: A Review” has been selected in International journal of information technologies and computer science and it will published information technologies and computer science and it will published on May June volume of journal. on May June volume of journal.
• Seminar: Published Research paper on “Symbolic Rules Extraction Seminar: Published Research paper on “Symbolic Rules Extraction From Trained Neural Networks” in two days UGC Sponsored From Trained Neural Networks” in two days UGC Sponsored National Seminar on Social Implications of Artificial Intelligence National Seminar on Social Implications of Artificial Intelligence organized in KMV College in Jalandhar. organized in KMV College in Jalandhar.
![Page 32: Symbolic Rules Extraction From Trained Neural Networks](https://reader033.vdocuments.net/reader033/viewer/2022051109/54825b01b079591f0c8b4774/html5/thumbnails/32.jpg)
26/5/201226/5/2012 Koushal KumarKoushal Kumar 3232
ReferencesReferences
• Olcay Boz AT & T Labs”Converting a trained neural network to Olcay Boz AT & T Labs”Converting a trained neural network to a decision tree DecText-Decision Tree Extractor.a decision tree DecText-Decision Tree Extractor.
• J.T Yao Dept of computer sci university of Regina “Knowledge J.T Yao Dept of computer sci university of Regina “Knowledge extracted from Trained neural networks whats next?extracted from Trained neural networks whats next?
• Simon Haykin ”Neural networks a Comprehensive foundation” Simon Haykin ”Neural networks a Comprehensive foundation” Pearson education(second edition).Pearson education(second edition).
• R. Davis, B.G. Buchanan, and E. Shortcliff, “Production Rules R. Davis, B.G. Buchanan, and E. Shortcliff, “Production Rules as aas a
• Representation for a Knowledge Based Consultation Progra”,Representation for a Knowledge Based Consultation Progra”,
• Artificial Intelligence, 1977, vol. 8(1), pp.15-45.Artificial Intelligence, 1977, vol. 8(1), pp.15-45.
• Knowledge Extraction from the Neural ‘Black Box’ in Knowledge Extraction from the Neural ‘Black Box’ in Ecological Monitoring Journal of Industrial and Systems Ecological Monitoring Journal of Industrial and Systems Engineering Vol. 3, No. 1, pp 38-55 Spring 2009Engineering Vol. 3, No. 1, pp 38-55 Spring 2009