Google makes PlaNet-AI that learns from open-source images

Google creates PlaNet, an AI agent that developed it, open-source. The technology makes use of reinforcement learning and learns with the help of images about the world. Google unveiled the AI in collaboration with DeepMind and wants to use the software to make plans.

For this reason, Google has christened the new artificial intelligence the Deep Planning Network (PlaNet). PlaNet learns a world model using the input it receives from images. It then uses that information to make plans. According to Google, the software is very efficient in solving image-related tasks. Google has made the source code for the artificial intelligence available on GitHub.

Dynamic models

Danjiar Hafner, one of the co-authors of an academic paper describing the architecture of PlaNet, explains how the software works. According to Hafner, PlaNet learns dynamic models based on the input it receives from images. With the help of these models, it then makes plans in order to gain new experiences.

Using those dynamic models, the software can get an understanding of various abstract concepts. Think for example of the speed of objects. The PlaNet agent learns this by predicting images. So it can take a bunch of images of a moving object and predict what the next frames will look like.

But the PlaNet software does not need to generate new images. Imagine, begins Hafner, the agent can imagine how the position of a ball and the distance of that ball from the goal changes by certain actions, but need not visualize that scenario. This allows us to compare 10,000 proposed actions with a large batch every time the agent chooses an action. We then carry out the action on the basis of the best inference and plan again at the next step.