Page - 125 - in The Future of Software Quality Assurance

Testing Artificial Intelligence 125 arenotprogrammedupfronttoperformspecifictasks, theyareempty.Thenodesare merely small calculators, processing parts they have been presented by top layers, returningacalculatedresult.Whentheneuralnetworkispresentedwithanexample in training it will systematically configure itself so the different layers and nodes will process parts and aspects of the input so the end result of all nodes will give the result that is given to the network (the label). Given two pictures, of a cat and of daddy, it will try different configuration in order to find the configuration that would determineone example to be a cat and the other as daddy. It would seek out thedifferencesso it’sconfigurationwouldcomeupwith therightclassificationnext time. In this way the neural network creates models of the labels: these reflect differences between a cat and daddy the neural network has identified based on the trainingdata. 2.3 Algorithm=Data+Code+Labels So what the system produces is an algorithm that consists of models derived from examples so it can classify and recognise input and assign these to labels. The algorithmis the productof the neural network but based stronglyupon the training data(theexamples)andthegoals(thelabels).So thealgorithmisNOTthecode,but thecode+ trainingdata+ labels.Because thealgorithmcannotbe identified it can also not be fixed directly.Brain surgerywill not fix the child’sflaws in recognising acat. 2.4 FuzzyLogics andMathematics Although all the system does is calculating, produce numbers, these numbers will not produce a Boolean result: for example: “this is daddy” or: “this is a cat”. The result will be the summation of all calculated numbers from the nodes and layers, eachgivinganumberwhichexpresses theextent towhichcriteriahavebeenmetas pereachgivenlabel.Thiswillhardlyeverbe1onascaleof0–1.Next to that: itwill also produce the extent to which the example scores on the other labels. So a new picture presented to the system could score “cat-ness” as 0.87 and “daddy-ness”as 0.13.Theconclusionwouldbe that theexampleisacat,but it’snot100%acat,nor is it 0%daddy. So the end product of AI is a calculation, a probability and never a 100% certainty.