The Problem with Current Post-hoc Machine Learning Interpretations and how Identifying Low-dimensional Structures can help

In the last decade machine learning algorithms have shown unprecedented accuracy in a variety of applications and tasks. However, current state-of-the-art machine learning algorithms are black-box models. As such, they make it hard if not impossible to understand the relationship between predictors and response.

In this talk I will survey and showcase general problems with post-hoc interpretations of machine learning algorithms. Local explanations need to be extrapolated to infer causal relationships which in turn makes the results arbitrary. Global explanations approximate the black box model by simple structures, like additive models. But a global approximation is just an approximation. In particular either the approximation is bad or, if the approximation is good, a simpler model should have been preferred in the first place.

Moving forward, I will present an alternative solution in the case where predictions are the composition of low dimensional structures by proposing local and global explanations that are exact. Examples of machine learning predictors that are compositions of low dimensional structures are gradient boosting machines and random planted forest.

Find the Q&A here: Q&A on' Machine Learning and its Opportunities'