Training AI Models to Collaborate with Humans: Improving Accuracy and Trust
Artificial intelligence (AI) models have proven to be highly effective in detecting patterns in images, often outperforming human observers. However, there are instances where human judgment still surpasses AI capabilities. To address this challenge, researchers at MIT and the MIT-IBM Watson AI Lab have developed a customized onboarding process that enables users to determine when to trust and collaborate with an AI assistant.
The training method involves identifying situations where the AI model’s advice may be unreliable, despite the radiologist’s inclination to trust it. By automatically learning rules for collaboration and describing them in natural language, the system guides the radiologist in making informed decisions. During the onboarding process, the radiologist engages in training exercises based on these rules, receiving feedback on her performance as well as the AI’s performance.
The results of the study indicate that this onboarding procedure led to a 5 percent improvement in accuracy when humans and AI collaborated on image prediction tasks. Notably, simply informing the user when to trust the AI without proper training resulted in poorer performance. The researchers emphasize that their automated system can be adapted to various tasks, making it applicable in domains such as social media content moderation, writing, and programming.
The absence of training materials and tutorials for AI tools is a common issue faced by users. The researchers aim to address this problem from both methodological and behavioral perspectives. They envision onboarding as an essential component of training for medical professionals, suggesting that doctors using AI for treatment decisions should undergo similar training. This approach could potentially revolutionize medical education and clinical trial design.
Unlike existing onboarding methods that rely on human-produced training materials for specific use cases, this system learns automatically from data. It collects data on human-AI collaboration for a particular task, embedding it into a latent space where similar data points are grouped together. Algorithms then identify regions in this space where the human collaborates incorrectly with the AI, capturing instances where trust in the AI’s prediction was misplaced. These regions are described as rules using natural language, forming the basis for training exercises.
The researchers tested their system on two tasks: detecting traffic lights in blurry images and answering multiple-choice questions from various domains. Users were divided into different groups, some of which underwent the researchers’ onboarding procedure while others followed a baseline onboarding process. The results showed that only the researchers’ onboarding procedure significantly improved accuracy without slowing down users. However, for the question-answering task, onboarding was less effective due to explanations provided by the AI model.
Providing recommendations alone without onboarding had a detrimental effect, leading to decreased performance and longer prediction times. Users became confused and struggled when solely relying on recommendations. On the other hand, onboarding had limitations related to the availability of data. If there is insufficient data, the effectiveness of the onboarding stage is compromised.
Moving forward, the researchers plan to conduct larger studies to evaluate the short- and long-term effects of onboarding. They also aim to leverage unlabeled data and develop methods to reduce the number of regions without omitting crucial examples.
Experts in the field recognize the significance of this work in improving human-AI interactions. The ability to identify situations where AI is trustworthy and effectively communicate this information to users is crucial for fostering better collaboration between humans and AI systems.
Funding for this research is provided, in part, by the MIT-IBM Watson AI Lab.Kindly read our copyright disclaimer here: https://cere-sync.com/dmca-copyrights-disclaimer/