The platform runs an environment that’s essentially a first-person video game, using it to train artificial intelligence (AI) agents. These “levels,” as DeepMind calls the AI agents (keeping with the game theme) can be created using off-the-shelf editing tools and are shared on GitHub, which is also where the platform itself will reside starting later this week.
DeepMind’s goal is to “solve” intelligence. That means going way beyond machine learning, which requires giving a computer thousands of examples and counter-examples. DeepMind is chasing the computer science discipline of deep learning, where the computer is able to educate itself.
The group’s best known achievement was AlphaGo, the AI agent that could beat a human at Go. DeepMind Lab is a different project and was not used to train AlphaGo.