Researchers from the University of Geneva are bridging a gap in the capabilities of artificial intelligence (AI), enabling one AI to learn tasks and communicate them to another AI for replication. The ability to grasp and communicate about new tasks represents a leap forward for AI, replicating a capacity inherent in human communication and consciousness. The findings of the study have been published in Nature Neuroscience.
This project was led by Alexandre Pouget, a professor at the UNIGE Faculty of Medicine, and his team. The research delves into natural language processing (NLP), a subset of AI that focuses on equipping machines to understand human language and respond intelligently.
Pouget explained the limitations of current AI capabilities. While AI can integrate linguistic information to produce text or images, translating a verbal or written instruction into a sensorimotor action or explaining it to another AI for replication has thus far proven elusive.
To tackle this, the team enhanced an artificial neural network (ANN), S-Bert, and connected it to a smaller network designed to simulate the language perception and production areas of the human brain — the Wernicke and Broca areas. Through training, this network can execute tasks based solely on written English instructions, then linguistically communicate these tasks to another network.
The range of tasks includes simple directives like pointing to a specific location and more complex instructions requiring the identification of subtle differences between visual stimuli. According to Ph.D. student Reidar Riveland, the integrated model created for the study connects the language-understanding S-Bert, which has 300 million neurons, with a simpler network comprising a few thousand neurons.
The AI system managed to both understand and execute instructions, performing unseen tasks correctly 83% of the time when given only linguistic instructions. Furthermore, the system was able to generate task descriptions that enabled a second AI to comprehend and replicate them at a similar success rate.
These advancements broaden the potential for AI to learn and communicate tasks linguistically, opening new avenues in robotics by integrating linguistic understanding with sensorimotor functions. This means AIs could be taught to understand and carry out instructions such as grabbing an item off a shelf or moving in a specific direction.
According to the researchers, the network created for this study is very small, paving the way for the development of much more complex networks integrated into humanoid robots capable of understanding not only human instructions but also the instructions of other AI. With increased investments flowing into AI and robotics, the possibility of intelligent humanoid robots may no longer be a futuristic concept but a looming reality.
