DeepMind Technologies Ltd, Alphabet Inc.’s artificial intelligence research group, is sharing more of its work with the world.

The division today open-sourced a collection of “key algorithmic components” derived from what it described as some of its most successful AI initiatives. The library, named TRFL  and pronounced “truffle,” aims to aid researchers working on projects that employ reinforcement learning, a popular way to train neural networks.

Reinforcement learning is essentially a method of using trial and error to hone an algorithm’s accuracy. Whenever an AI makes a correct decision in a training environment, it receives a kind of virtual reward, which keeps the neural network’s development on the right track.

TRFL includes implementations of mathematical operations commonly used by algorithms that rely on reinforcement learning. They’re paired with components that DeepMind said can carry out more “cutting-edge” calculations, as well as various other building blocks including tools for ensuring AI training sessions go smoothly.

The entire collection is built to run on the popular TensorFlow deep learning engine created and open-sourced by Alphabet’s Google LLC. Researchers can interact with TRFL through an application programming interface that, according to DeepMind, makes it relatively straightforward to combine the components inside with technologies and concepts from other sources.

The group’s goal with the library extends far beyond just simplifying individual reinforcement learning projects. TRFL is part of a broader effort by DeepMind to create common building blocks for AI researchers to draw on, an effort that has seen the unit open-source other internal software over the years.

The basic idea is that if the same components are reused across across projects, researchers will have an easier time reproducing their colleagues’ work. The ability to more closely replicate the parameters of AI projects would in turn boost debugging efforts. According to DeepMind, outside reviewers often play a big role in identifying flaws affecting neural networks or associated components.  

“These parts tend to interact in subtle ways (often not well-documented in papers, as highlighted by Henderson and colleagues), thus making it difficult to identify bugs in such large computational graphs,” the group said. “A recent blog post by OpenAI highlighted this issue by analysing some of the most popular open-source implementations of reinforcement learning agents and finding that six out of 10 had subtle bugs found by a community member and confirmed by the author.”

TRFL is available on GitHub. It adds another item to the long list of open-source projects that DeepMind and Alphabet have released in recent years.

Photo: DeepMind