All Times EDT
The frenzy of anyone being able to create AI algorithms is priming the future for a reality check – that of decay once the algorithm is placed in the field. The lack of risk assessment and multiple error rate comparisons during model development as well as robust post-release analytic reviews can lead to costly surprises when accuracy takes a dive for the worst. Unforeseen or out-of-date changes to labels, data, customer or internal architecture, tolerances, and test sets can wreak havoc to customers and developers alike, causing potential loss to revenue as well as to valuable brand image. Contributing to decay are those “unknown unknowns” such as Adversarial Machine Learning attacks (“Model Hacking”), that suddenly impact a seemingly resilient model. This talk will describe the challenges of decay and corrective actions to enable continuous improvement in analytics, their processes, and the products they support.