A Distinguished Lecture with Mireille Hildebrandt Event

Interpretation and Prediction: MT in ML

This lecture is not about machine translation (MT) in machine learning (ML), though admittedly ML requires lots of translation (pun intended). Instead, I will discuss the cross-roads of the ‘interpretability problem’ in ML and its predictive salience as a problem on the brink of magical thinking (MT) and speculative exploration in ML. Web Science has been said to target two magics (things we don’t understand, yet): the influence of microscopic online interaction on macroscopic real world events, and the detection of new microscopic designs that could have positive macroscopic effects. I propose to define magic thinking as a mistaken belief in causal influences and/or predictive accuracy, based on illusionary wishful thinking about the nature of our world. This is directly related to the so-called interpretability problem in ML and the urge to favour math-inspired prediction above causal interpretation. My point will be that interpretation concerning social interaction concerns both meaning and causality, requiring us to denounce magical beliefs in predictive accuracy. Seeking microscopic design for positive real world effects calls for (1) designs that foster the multi-interpretability of human interaction, and (2) designs that reduce the manipulability of ‘user behaviour’ based on quantified models of influencers. This should result in a world wide web that affords speculative exploration as well as explainability, introducing new ways to confront the performance metrics of ML with the performativity of human society.