Researchers from MIT and the MIT-IBM Watson AI Lab have developed an automated training system that can guide users on when and how to collaborate with AI models effectively. The system, designed to adapt to multiple tasks, does this by training users using data from the interaction between the human and AI for a specific task. The team asserted that this training mechanism enhances the accuracy of results by guiding users on when to trust AI assistants’ predictions and when to proceed with their judgement.
Artificial Intelligence (AI) models, particularly those used in medical image prediction tasks such as radiology, often outperform humans in identifying patterns but are not infallible. The MIT-IBM research team aimed to address significant gaps in training with their onboarding procedure, testing it in tasks that included identifying a traffic light through blurry images and answering multiple-choice questions across different domains. Results from these exercises showed that only the researchers’ onboarding procedure enhanced users’ overall accuracy performance by approximately 5%.
The automated onboarding method works by collecting data on the human-AI interaction and progressively identifying regions of failure in accuracy, such as where the human trusts an incorrect prediction by the AI or vice versa. The system then describes identified regions as rules, utilizing natural language and explaining examples that may create contrasts. Consequently, these rules become part of the training. Users answer questions based on these rules and receive feedback that guides their collaboration with the AI to improve accuracy.
According to the researchers, the onboarding process poses a significant value proposition to medical professionals and other fields such as content moderation, writing, and programming. However, they added that the extent to which the onboarding process can be effective depends on the availability of data, necessitating further studies to evaluate the long-term impacts of the procedure.
The researchers’ objective is a broader realization of the potential of AI, enabling users to comprehend when AI suggestions would be reliable and when they may be liable to errors.
