AI developed by DeepMind demonstrates social learning ability

A research team from Google's DeepMind project has developed an AI system that demonstrates social learning abilities. The team describes how they developed an AI application that allows it to learn new skills by replicating the actions of "experts" implanted in a virtual world.

Most AI systems, such as ChatGPT, acquire knowledge by learning from large amounts of data. However, industry experts point out that this approach is not very efficient. Therefore, many researchers in this field continue to search for other methods to teach AI systems.

One of the most popular methods used by researchers is to try to mimic the process of human learning. Like traditional AI applications, humans learn by being exposed to known elements in the environment and following examples of others who know what they are doing. However, unlike AI applications, humans can master things without a large number of examples. For example, a child can learn to play hopscotch after observing others for a few minutes. In this new work, the research team attempts to replicate this process using AI confined to a virtual world.

The team's work begins by creating a virtual world called GoalCycle3D, consisting of uneven terrain with various obstacles and colorful spheres placed on it. They then add AI agents that are supposed to navigate the virtual world by avoiding obstacles and passing through the spheres. These agents are equipped with learning modules but have no other information about the world they will inhabit. They acquire knowledge of how to move forward through reinforcement learning.

In order to make the agents learn, they are rewarded and allowed to autonomously explore routes through multiple similar virtual worlds over and over again. By doing so, the agents are able to navigate the virtual world to reach the desired destination. The researchers then add another feature to the virtual world - expert agents who already know the optimal path from one place to another without encountering obstacles. In the new scenario, the non-expert agents quickly learn from the experts how to reach the desired destination in the fastest way.

While observing the learning process of the agents, the researchers found that with the presence of expert agents, they learn faster and are able to navigate other new similar virtual worlds better by imitating what they learned from the experts in previous experiments. Even without the presence of experts, they are able to apply these skills (with the help of memory modules), which the researchers claim is an example of social learning.