Researchers demonstrate AI systems can learn from implicit human feedback
Researchers at Google Brain and the Massachusetts Institute of Technology have been working on demonstrating that artificial intelligence (AI) agents can learn from implicit social feedback from humans, and use that feedback to improve themselves. In their experiment, the researchers built an application that showed people samples of drawings generated by an AI system, and recorded their facial expressions. The recorded images were fed into a facial-expression detection network and used to compute emotions such as amusement, contentment, concentration, and sadness. The resulting data was then used to train the AI system to produce better drawings. The experiment demonstrated that ‘implicit social feedback in the form of facial expressions not only can reflect user preference, but also can significantly improve the performance of a deep learning model’. The researchers believe that making AI models improve in quality as a result of learning from implicit human feedback is a significant step towards improving AI safety: ‘an AI agent motivated by satisfaction expressed by humans will be less likely to take actions against human interest’.