AI masters social learning in real-time without pre-collected human data: DeepMind study

tech.hindustantimes.com:

Human intelligence has long relied on cultural evolution and social learning, where knowledge is passed down through imitation of actions and behaviors down through the ages. The question arises: Can artificial intelligence (AI) replicate these social learning skills in real time? Traditionally, AI has employed imitation learning, observing humans and attempting to mimic their actions. However, this method often requires numerous examples and extensive exposure to data. In a groundbreaking study, DeepMind researchers claim that AI agents can exhibit social learning skills in real time without relying on pre-collected human data.

The study focused on observational learning, specifically the imitation of body movements in novel contexts. DeepMind conducted experiments in GoalCycle3D, a simulated environment with varied terrain, footpaths, and obstacles that AI agents navigated. To facilitate learning, reinforcement learning was employed, offering rewards for behaviors leading to the desired result. Also read: Talk to celebrities via AI chatbot! Know how to create your own character on Character.ai

In the next phase, expert agents, either hard-coded or human-controlled, were introduced. The AI agents quickly learned that mimicking these experts was the most effective way to navigate. Notably, the AI not only learned faster when imitating experts but also applied this knowledge to navigate other virtual paths. Furthermore, the AI agents demonstrated the ability to use their newfound skills even in the absence of experts, showcasing a form of social learning.

While acknowledging the need for further research, the authors believe their method could pave the way for cultural evolution to play a crucial role in the development of