Probabilistic, Not Deterministic: Why Testing ML is Different

Why can’t we test ML systems as we test software? Have you ever wondered?

Machine learning systems have characteristics distinct from traditional software that demand fundamentally different testing approaches. In this talk, we make the case for why testing ML requires moving beyond established software testing paradigms.

We first outline key differences including probabilistic outputs, black box internals, dependence on data, and emergent behaviours. We then discuss limitations of traditional testing techniques like requirements-based, unit testing, UI validation when applied to ML systems.

Core testing challenges with ML systems associated with continuous evolution, feedback loops, subjectivity in quality evaluation are highlighted. We present examples of failures and risks specific to ML like bias, adversarial samples, and model decay.

The talk covers recommendations on adapting techniques like metamorphic testing, fuzzing, simulated environments and online testing for ML systems. We advocate expanding the definition of “code coverage” to account for data distribution, model robustness and fairness defects.

We will share critical insights into why ML necessitates rethinking testing approaches with a shift left mindset and emphasis on behaviour testing in realistic environments. The talk aims to equip practitioners with perspectives to deliver comprehensive and rigorous testing for ML powered software.