In a groundbreaking experiment presented at the Neural Information Processing Systems (NeurIPS) conference, researchers from the University of British Columbia reveal that artificial intelligence (AI) equipped with an inner monologue demonstrates significantly improved learning capabilities. The study, conducted by computer scientists Shengran Hu and Jeff Clune, explores the integration of language with actions to enhance an AI program's proficiency in mastering complex tasks.
The researchers designed an AI agent tasked with completing missions in a virtual 2D world. This agent, employing two neural network components, connected language, actions, and observations to execute intricate tasks in a 20-by-20 grid. The innovation lies in the agent's ability to generate thoughts, such as "open blue door to explore" or "go to purple box," aligning language with actions and significantly improving performance.
Through a meticulous training process using a data set of missions completed by a specialized bot, the AI agent demonstrated an 80% success rate on complex missions when trained to imitate both actions and thoughts. In contrast, an agent trained solely on imitating actions achieved success only 30% of the time. This approach, termed "thought cloning," not only enhances AI performance but also provides transparency, allowing human operators to understand the AI's thought processes.
Furthermore, the researchers explored the potential for "precrime intervention," illustrating the ability to halt AI agents before executing prohibited actions. By implementing rules triggered by the AI's thoughts, the researchers effectively prevented actions like touching a red item without retraining the model.
Looking ahead, the researchers envision incorporating an inner monologue component into pretrained models, such as OpenAI's GPT-4 Vision, enabling AI to glean insights from vast information sources like YouTube videos. This innovative approach opens new possibilities for AI to learn practical skills by emulating human thought processes.
As the world witnesses this transformative leap in AI learning, the integration of inner monologue emerges as a powerful tool, bridging the gap between artificial and human intelligence.
More: https://www.science.org/content/article/artificial-intelligence-may-benefit-talking-itself
