Google DeepMind introduces AI with social learning capabilities.
Scientists from Google DeepMind have created an artificial intelligence that can learn new skills just by observing human actions. Usually, learning algorithms to simulate human actions requires hundreds or thousands of examples, but this AI is able to learn new skills on the fly.
The human ability to learn from each other quickly and effectively is one of the most important aspects of social learning, also known as cultural transmission. Researchers have long tried to replicate this process in machines, using an approach in which AI watches a person perform a task and then tries to mimic its behavior.
DeepMind researchers have taken a step forward by training AI agents to navigate the virtual world in real time, based on human actions. "Our agents successfully mimic a human in real time in new contexts without using pre-collected human data," says the study, published in Nature Communications.
For training agents, a special GoalCycle3D simulator was created, which generates an almost infinite number of different environments based on the rules of operation and variability of the simulation. In each environment, AI agents who look like small blobs must overcome uneven terrain and various obstacles to pass through a series of colored spheres in a specific order.
Agents are trained to navigate using reinforcement. They receive a reward for passing through the spheres in the correct order and use this signal to improve their results. However, environments also have an expert agent that already knows the correct route.
Over the course of many training sessions, the AI not only learns the basics of how environments work, but also understands that the fastest way to solve a problem is to simulate an expert. To make sure that the agents really learn to simulate, and not just remember routes, the team trained them in one environment, and then tested them in another.
However, it can be difficult to translate the approach into more practical areas. The limitation is that during all training runs, the expert agent was supervised by one person, which makes it difficult to understand whether agents can learn from different people. In addition, the ability to randomly change the learning environment is difficult to replicate in the real world.
However, progress in AI social learning is an important step. If we want to live in a world with intelligent machines, effective and intuitive ways to transfer our experience and knowledge to them will be extremely important.
Scientists from Google DeepMind have created an artificial intelligence that can learn new skills just by observing human actions. Usually, learning algorithms to simulate human actions requires hundreds or thousands of examples, but this AI is able to learn new skills on the fly.
The human ability to learn from each other quickly and effectively is one of the most important aspects of social learning, also known as cultural transmission. Researchers have long tried to replicate this process in machines, using an approach in which AI watches a person perform a task and then tries to mimic its behavior.
DeepMind researchers have taken a step forward by training AI agents to navigate the virtual world in real time, based on human actions. "Our agents successfully mimic a human in real time in new contexts without using pre-collected human data," says the study, published in Nature Communications.
For training agents, a special GoalCycle3D simulator was created, which generates an almost infinite number of different environments based on the rules of operation and variability of the simulation. In each environment, AI agents who look like small blobs must overcome uneven terrain and various obstacles to pass through a series of colored spheres in a specific order.
Agents are trained to navigate using reinforcement. They receive a reward for passing through the spheres in the correct order and use this signal to improve their results. However, environments also have an expert agent that already knows the correct route.
Over the course of many training sessions, the AI not only learns the basics of how environments work, but also understands that the fastest way to solve a problem is to simulate an expert. To make sure that the agents really learn to simulate, and not just remember routes, the team trained them in one environment, and then tested them in another.
However, it can be difficult to translate the approach into more practical areas. The limitation is that during all training runs, the expert agent was supervised by one person, which makes it difficult to understand whether agents can learn from different people. In addition, the ability to randomly change the learning environment is difficult to replicate in the real world.
However, progress in AI social learning is an important step. If we want to live in a world with intelligent machines, effective and intuitive ways to transfer our experience and knowledge to them will be extremely important.
