Seite - 127 - in The Future of Software Quality Assurance
Bild der Seite - 127 -
Text der Seite - 127 -
Testing Artificial Intelligence 127
considered as a cause-effect analysis of unwanted outcome. This could give clues
foroptimising thesystem.Forexample:under-fittedsystemsmost likelyneedmore
diverse trainingdata,over-fittedsystemsstreamliningof labels.
3.1 Bias
The main risks with AI are types of “bias”. In human intelligence we would call
this prejudice, reductionism or indecisiveness. Because of limits in training data
and concepts, we see things too simple (reductionism) or only from one point of
view (prejudice). A high granularity in concepts could mean that the system can’t
generaliseenough,making theoutcomeisuseless (indecisiveness).
Significant typesofpossiblebias in AIarediscussednext.
3.1.1 SelectionBias
If the training data selection misses important elements from the real world, this
could lead to selection bias. Compared to the real results, the polls for the last
European elections predicted much higher wins for the Eurosceptic parties in the
Netherlands than they did in the real election. The polls did not filter on whether
people were really going to vote. Eurosceptics provedmore likely not to vote than
othervoters.
3.1.2 ConfirmationBias
Eagerness to verify an hypothesis heavily believed or invested in can lead to
selecting or over-weighing data confirming the thesis over possible falsifications.
Scientists, politicians and product developers could be susceptible to this kind of
bias, even with the best of intentions. A medical aid organisation exaggerated a
possible food crises by showing rising death numbers but not numbers of death
unrelated to famineand theoverallpopulationnumber inorder to raisemorefunds.
3.1.3 Under-fitting
Training data lacking diversity causes under-fitting. The learning process will not
becapable todeterminecriticaldiscriminatingcriteria.Software thatwas trained to
recognisewolvesfromdogs, identifieda huskyasa wolfbecause it hadnot learned
thatdogscanalsobeseen insnow.Whatwouldhappenifweonlygetdrugs-related
newsmessages in theNetherlands?
The Future of Software Quality Assurance
- Titel
- The Future of Software Quality Assurance
- Autor
- Stephan Goericke
- Verlag
- Springer Nature Switzerland AG
- Ort
- Cham
- Datum
- 2020
- Sprache
- englisch
- Lizenz
- CC BY 4.0
- ISBN
- 978-3-030-29509-7
- Abmessungen
- 15.5 x 24.1 cm
- Seiten
- 276
- Kategorie
- Informatik