DeepMind's research reveals AI agents learning new skills by mimicking human behavior, a key step towards advanced artificial general intelligence.
Google's DeepMind has made a significant stride in AI development, showcasing that AI agents can mimic human behaviors to enhance their learning process. This breakthrough, led by Edward Hughes, a research engineer at DeepMind, presents a novel approach to AI learning, drawing parallels with social learning observed in humans and animals.
Traditionally, AI skill acquisition has relied heavily on supervised learning from extensive human demonstrations, a method both time-consuming and costly. The DeepMind team, however, explored a more efficient technique inspired by human learning methods. In their study, AI agents were placed in a simulated environment named GoalCycle3D, designed like a virtual playground with various paths and obstacles. Remarkably, these agents successfully learned from both human and AI mentors in navigating complex tasks, despite having no prior exposure to humans.
The findings, published in the Nature Communications journal, highlight the team's use of reinforcement learning. This approach enabled the AI agents to identify new experts, imitate their actions, and retain this knowledge efficiently. The study emphasizes the simplicity of the methods used to facilitate cultural transmission in AI, a step forward in the development of artificial general intelligence.
Looking beyond their current achievement, the researchers envision broader applications of their findings in AI. They propose that cultural evolution, defined as the development of skills over generations within a community, could be a key component in advancing AI capabilities. Future experiments might explore the accumulation of behaviors and skills across generations of humans and AIs in lab settings. Such research could provide valuable insights into the dynamics between mixed populations of humans and AI agents.
The DeepMind team is optimistic about the potential for cross-disciplinary collaboration between AI and cultural evolutionary psychology, anticipating that such interactions will yield fruitful advancements in both fields.
