Study: Curious robots mimic how children can learn to understand language (www.oist.jp)

Researchers at the Okinawa Institute of Science and Technology (OIST) have developed a virtual robot using a brain-inspired neural network that mimics how children learn language, driven by a unique trait: curiosity. By immersing these robots in a diverse linguistic environment filled with verb-adjective-object combinations, the study revealed that curiosity enhances language comprehension, leading robots to understand language in half the time compared to those that lack it. This finding, published in Science Advances, provides insights into the mechanisms behind children’s rapid language acquisition, suggesting that a rich and varied language exposure, coupled with curiosity, is pivotal for learning. The robots leverage a Predictive coding-inspired Variational Recurrent Neural Network (PV-RNN), which balances the desire for accuracy with the impulse to explore, thereby facilitating learning through trial and error. The study also observed behaviors paralleling those of toddlers, such as the U-shaped exception-handling performance curve, where initial success in language usage can decline as rules are learned, but rebounds as mastery increases. This transparent model not only sheds light on human cognitive processes but also establishes a new framework for studying language acquisition in AI, pointing to significant implications for both AI research and our understanding of child language development.