Richard Self LLM
Software Testing Imperatives for AI and Machine Learning systems
As an industry, we feel that we understand the processes involved in software testing of algorithmic systems, where the specifications determine exactly what the system will do, mostly in terms of business processes.
However, when we consider the self-learning systems that typify AI and Machine Learning, this is no longer sufficient. It is possible to use standard software testing approaches to verify and validate that the software meets the specifications but not that software then learns the correct and unbiased behaviour that is required.
These learning systems are defined to be able to find the patterns in the training data, whether with supervised or unsupervised training. In principle, these pattern finding systems will always find patterns in sufficiently large sets of data, many of which are entirely spurious. In addition, if the training data does not have sufficient diversity, that is it is biased, the systems will learn biased patterns.
This session will identify some of the critical factors that need to be addressed in order to mitigate some of the current problems with AI and Machine Learning systems.