Seite - 126 - in The Future of Software Quality Assurance

126 G.Numan 2.5 Development andCorrection Developmentofaneuralnetworkconsistsofdevelopinganeuralnetworkitself,but mostdevelopers takeaneuralnetworkoff theshelf.Next theyneed toconfigurethe neuralnetworkso it can receive the inputathandandconfigure labels, soexamples are linked to these. Finally the layers of the neural network can be parameterised: the calculated results can be weighted so certain results will have more impact on the end result thanothers.Thesearethemain tweakinginstrumentsdevelopershave. If thesystem isnotperformingsatisfactorilytheparameterscanbetweaked.Thisisnotafocussed bugfix, correctingonecaseof faultydecision. Parametrisation will influence the outcome, but each tweak will have impact on the overall performance. In AI there is massive “regression”: unwanted and unexpectedimpactonpartsof thesystem thatarenot intended tobechanged. Trainingdata and labels are also likely candidates for influencing the system. In certain issues with AI, such as underfitting, expanding the training data will very likely improve the system. Underfitting means the algorithm has a too simplistic viewof reality, forexamplewhenacat isonlyclassifiedasa furrycreature.Adding more examplesof a cat to the training data, showing more variety of species, races andbehaviour,couldhelp thesystem distinguishacat fromothercreaturesbetter. 2.6 Overall VersionEvaluationandMetrics Whenbugfixescannotbefocussedandeachtweakhasmassiveregression,massive regressiontestingisnecessary.Thequestion“didwefixthisbug?”becomesaminor issue. We want to know the overall behavioureach time we change something.We want to know what the overall performance of the system is compared to other versions. In that overall evaluation we need to take into account the output of AI: calculatedresultswhicharenoteither trueor false.Eachresult isagradeonascale. So the end results should be thoroughly compared, weighed and amalgamated so we can decide if a versionas a whole is better than anotherand we should use it or not.Theresultwill bemetricscalculating thevalueofoutputbasedonexpectations and their relative importance. 3 Risks in AI We’ll discuss themost important riskshere.These risksare typicalofAIandcould have serious impact on the quality of AI, it’s customers, users, people and even the world. These risks should be considered before starting testing, giving clues to where to put emphasis as a tester. When analysing test results the risks should be