Addison Shuppy

File explanations: where: Model - a trained convolutional neural network.

simpleNetwork.m: Creates a simple neural network with one hidden layer, three inputs, 3 hidden nodes, and 1 output node. Weights and biases are set manually.

generateExamples.m: Given a model. Returns training data for that model (trainingData.xls). This simulates the scenario where we are given a trained Neural Network and training examples for which the model is correct in 95% of examples.

knowledge_extraction.m: Given a model. Extracts logical rules for that model (rawRules1.csv) Uses an adaptation of the methodology of extraction in Tran, Garcez 2000.

ruleReduce.m: Given rules from knowlege_extraction.m, Simplifies the number of rules by removing duplicates and subsuming. I.e. c <- a, ~b; c <- a, b; reduces to c <- a.

knowledge_extraction_b.m: Given a model. Extracts confidence rules for that model (confidenceRules1.csv) Extracts rules from a network using the algorithm in Tran, Garcez 2018, modified to apply to convolutional neural networks instead of Deep Belief Networks.

logicalNetwork.m: Incomplete.

confidenceRuleEncoding.m: Given a model, confidence rules, and training data. Trains a new Logical Hybrid Neural Network.

logicalRuleConvert.m: Given logic rules from knowledge_extraction.m or ruleReduce.m, Revises rules to their true propositional logic form.

truthValueGenerator.m: Given rules from ruleReduce.m (reducedRules1.xlsx), calculates the output for all possible input values.