Page - 130 - in The Future of Software Quality Assurance

130 G.Numan Traditionally requirementsand specificationsare determinedupfrontand testers receive themreadytobeusedat thestart. InAI, requirementsandspecificationsare too diverse and dynamic to expect them to be determined at the start completely and once and for all. Product owners and business consultants should deliver requirements,but testers need to take initiative to get the requirements in the form, granularityandactuality that theyneed. The challengeswith testing AI and their accessory measures from start to finish arediscussednext. 4.1 Review of theNeural Network, TrainingData andLabelling Static testingcandetectflawsor riskyareasearly. Thechoicefortheneuralnetworkoritssetupcanbeassessed:isitfitforpurpose? What are the alternatives? For this review a broad knowledge is required of all possibleneuralnetworksand their specificqualitiesandshortcomings. The trainingdataand labels canbe reviewedandassessed for risk sensitivity: 1. Does the data reflect real-life data sources, users, perspectives, values well enough? Could there be relevant data sources that have been overlooked? Findingsmight indicate selectionbias, confirmationbiasorunder-fitting. 2. Are data sources and data types equally divided? How many representatives do various types, groups have compared to one another? Findings might indicate under-fitting,selectionbias, confirmationbiasoroutliers. 3. Are the labels a fair representation of real-life groups or types of data? Do the labels match real-life situations or patterns that the system should analyse? Findingsmight indicateover-fitting,under-fittingorconfoundingvariables. 4. Is the data currentenough?What is the desired refresh rate and is thismatched? Are thereevents in the real world that arenot reflectedwell enoughin the data? 4.2 IdentifyingUsers Theownerof thesystemisnot theonlyvaluableperspective!AI-systemslikesearch systems are an important part of the world of its users but also of those that are “labelled” by it. The quality of an AI-system can have moral, social and political dimensionsand implicationsso these need to be taken intoaccount. The users of AI are often diverse and hard to know. They are not a fixed set of trained users, all gathered in a room and manageable in their behaviour and expectations.Theycouldbe thewholeworld, like in thecaseofasearchengine:an AmericantouristvisitingAmsterdamoranexperiencedart lover in thefieldathand have very different needs and expectationswhen searching for “Girl with pearl” in