When you start learning to drive, your instructor gives you instant tips on every move. They might even grab the wheel and yell "Brake!" to keep you safe. This direct teaching builds your driving skills over time. However, teaching artificial intelligence (AI) to drive or perform tasks is different. AI usually learns from huge amounts of data, not from real-time advice like you get from a driving coach.
Now, researchers from Duke University and the Army Research Laboratory have created GUIDE, a new AI teaching method. GUIDE helps AI learn like a human would, with real-time feedback. They will show GUIDE at the Conference on Neural Information Processing Systems (NeurIPS 2024) conference in Vancouver. Boyuan Chen, a professor at Duke, explains that AI often struggles with tasks needing quick decisions based on little information. GUIDE changes this by using ongoing human feedback instead of big data sets.
Fast human-assisted AI training
In GUIDE, humans watch what the AI does and give detailed instructions, much like a driving coach. In one study, they used GUIDE to teach AI to play hide-and-seek. Here, a human trainer guides an AI-controlled red player to find a green one on a playing field with obstacles. The human provides feedback using a scale, not just saying if it's good or bad, allowing for more precise guidance.
This study, involving 50 people with no special training, showed that just ten minutes of feedback could boost AI performance by 30%. After human coaching, the AI could continue learning from a simulated coach, showing that human guidance can be brief yet effective. Chen points out that even if the coach isn't the best at the task, they can still teach well, just like many sports coaches.
The researchers found that people's different skills, like quick thinking or understanding space, affect how well they can guide AI. This discovery opens up new ways to train both humans and AI better.
GUIDE could make AI more adaptable and intuitive, working well even with limited information, thus making AI more useful in real-life situations where things change quickly.