Personalized machine learning

Personalizing a machine learning system can’t be accomplished by developers alone—the end user needs to be involved. Big data can help learning systems to generalize well, but because no two humans are exactly alike, we can’t rely on generalization alone. Perhaps the user’s task is critical (e.g., a smart home environment to help monitor an aging relative), or novel (i.e., there is no existing set of training data), or the system is generalizing in a manner contrary to the user’s goal. Even if the system is mostly correct (based on the user’s notion of correctness, not that of the developers), people should still have the ability to improve the system’s predictions. In any of these cases, the end user needs to be able to communicate with the learning system to align its internal logic with his or her own reasoning.