Facebook creates AI toolkit Horizon for reinforcement learning models open source

Facebook has made an important part of its Artificial Intelligence (AI) software open source. This allows developers to create reinforcement learning models in their software applications, reports Silicon Angle.

Developers can now use the Horizon Toolkit to build applications that can learn to perform specific computing tasks through trial and error. Facebook itself has already used Horizon for various tasks, including teaching its systems to predict which notifications users are likely to respond to.

Horizon, for example, helps to understand that users are more likely to respond to a “like” from his mother than from dozens of others who also responded to the message. As a result, mothers are displayed more prominently to ensure that a user responds. Horizon was also used to give personal suggestions via assistant M in Messenger.

Reinforcement learning is an AI component that, among other things, uses simulated environments to teach programs how to perform certain tasks. For example, Facebook uses this to determine whether to stream high or low quality videos to users, depending on the strength of their mobile connection or their location.

Milestone

In this system, computers are rewarded or punished on the basis of the outcome of their actions. In the case of the notifications on Facebook, the engineers gave the systems rewards if a notification sent to a user resulted in a response. If there was no response, the system was punished. In this way, the systems learned which notifications should be given priority in order to get more reactions in the form of likes or reactions.

Holger Mueller, analyst at Constellation Research, states that making Horizon open source is a milestone, because reinforcement learning on a large scale is difficult to achieve for companies. With this decision, the platform will probably be adopted more often and more attention will be paid to it. However, challenges remain because Horizon runs on the PyTorch machine learning framework, which has difficulty keeping up with Google’s more popular TensorFlow.