Our Research

Recent successes in AI research may give the impression that society will soon benefit from autonomous systems replacing human labor. Behind the scenes, however, tedious manual tasks will remain an integral requirement of machine learning systems for the foreseeable future, be it in the form of monitoring outputs for failures, explaining autonomous decisions, or labeling vast amounts of training data.

The field of Interactive Machine Learning tackles these challenges by setting aside the paradigm of fully autonomous systems and instead placing the human at the center of AI application: Only the design of human and machine as a symbiotic system from the ground up allows to harness manual labor in the most efficient way and thus its reduction to a bare minimum.

Following this human-centered AI perspective, we focus on 3 impact areas:

Active Learning

To alleviate the burden of manually labeling vast amounts of training data, we enable human-in-the-loop training by transforming state-of-the-art machine learning models into interactive tools.

Explanatory AI

Comprehending a model’s predictions lies at the heart of human-machine interaction. We aim to extract the causal reasoning behind decisions and render insights accessible to domain experts.

Failure Detection

One aspect of symbiotic decision making strives to alleviate the burden of monitoring predictions from imperfect models. To this end, we research the ability of systems to detect and flag their own failures.