Page - 124 - in The Future of Software Quality Assurance

124 G.Numan strengthenedprejudicedopinionsbecause it wasoptimised to rewardclick-success. Negative informationshowedup in searchresults increasingly. Every software developer and customer of AI struggles with these doubts and risks.What isabugincaseofAIandhowtofixit?Howtobecertainthat thesystem does the right thing with a great variety of input and users? How to get the right levelofconfidence?Aretheresults fair toall concerned?Arecurrentdevelopments, opinionsandvalues reflected in thealgorithm? WhatarethebiggestriskswithAIandhowtodealwith themfromatestingpoint ofview? 2 An Introductionto AI for Testers This chapter is a short introduction to AI and an analysis of aspects relevant to testing. 2.1 AIIs BlackBoxDevelopment In AI, the algorithm, the behaviourof the system in termsof criteria, decisionsand actions, are not explicitly engraved in the code. In non-AI development the code directlyexpresses thealgorithm. InAI thealgorithmis theproductof trainingdata, parameterisation, labelsandchoiceof theneuralnetwork.But thealgorithmcannot be found in the code. The code, the neural network, is just a, be it very essential, part of a system which produces the algorithm by training. This is the essence of machine learning. 2.2 MachineLearningandNeuralNetworks There is a strong analogy between machine learning and human learning. Take for example a child who learns to use a concept for the first time. The child has been told that the hairy creature it cuddles is a “cat”. Now the child sets its own neural network to work. The concept of the cat is compared to objects which aren’t cats, suchas“daddy”.Theneuralworks is findingways toconfigure itself in suchaway that had it seen the cat, it would classify it as a cat and not as daddy. It does so by finding differences, criteria, such as fur, whiskers, four legs, etc. But we do not knowexactlywhat thesecriteriaare.Theymightalsobe“huntingmice”,“purring”, or “being white”. We cannot find the concept of a cat and it’s criteria inside the brain,norcanwe correct it directly in thebrain. A neuralnetworkconsistsofmanyblocksofcode (“nodes”)whicharearranged in layers. Each layer of nodes is connected to its top and bottom layers. The nodes