Ensembling diverse models catches errors that single models miss.

Different artificial intelligence models can achieve the exact same overall accuracy while still making conflicting predictions on the very same examples. This conflict is called predictive multiplicity. When these models are combined into an ensemble, understanding this problem becomes difficult. To solve this, researchers designed a way to audit decision systems and flag questionable cases for human review. Their method checks how confident the group of models is, known as the ensemble marginensemble marginA measure of how confident a group of models is in their combined decision., along with how sensitive each individual model is to small changes, called local prediction variabilitylocal prediction variabilityHow much individual models' predictions vary around a specific input point.. If the score shows inconsistency, the system sends the instance to a person. The researchers found that auditing an ensemble greatly reduces the risk of wrong predictions slipping past unnoticed compared to checking just one model. It does this with only a moderate increase in flagged cases. Furthermore, small ensembles closely match the auditing performance of checking every possible accurate model.

Why it matters

When AI systems make decisions, some mistakes go unnoticed if only one model is checked. This method helps catch those errors by using a group of models, making audits more reliable for people affected by AI choices.

Questions answered