2 min

Tags in this article

, , , ,

Google introduces the new open source framework SEED RL, which enables companies to reduce training costs for an artificial intelligence (AI) model. According to the developer, the costs would be reduced by up to 80 percent.

SEED RL is built on top of the Google machine learning platform TensorFlow. The framework uses graphics processing units (GPUs) and tensor processing units (TPUs) for model inference. The inference process is performed centrally by the learner component that trains the model.

The variables and status information of the model remain local, but observations are continuously sent to the learner component. Google indicates that the learner can be scaled to thousands of cores, for example up to 2,048 cores at Cloud TPUs. Thousands of machines can also be used for other training tasks.

Speed gain

To see how this works out in practice, Google carried out the Football Benchmarks. A reinforcement learning environment is used to get the hang of football. SEED RL was able to solve a Football task, training the model with 2.4 million frames per second with 64 Cloud TPU chips. According to Google, that’s 80 times faster than previous frameworks.

This improvement in speed reduces the cost of training, due to the fact that model training normally involves various experiments and tests.

Google has taken several steps in recent months to promote the development of AI. For example, it recently released a new version of photo-dataset Open Images, to offer extra possibilities for labelling objects.