[ad_1]
At the F8 developer conference, Facebook today launched Ax and BoTorch, two new open source AI tools.
BoTorch, which, as its name implies, is based on PyTorch, is a Bayesian optimization library. It is a fairly specialized tool. Ax is the most exciting launch because it is a versatile platform for managing, deploying and automating AI experiments. However, both tools are part of the same global work at Facebook, which focuses on what the company calls "adaptive experimentation." Indeed, Ax interfaces with BoTorch and, internally, Facebook has used both of these tools for tasks that range from Instagram optimization. back-end infrastructure to improve the response rates of user surveys.
BoTorch – and Bayesian optimization in general – is essentially aimed at making model optimization easier and faster for data scientists to get a ready-to-go model. As a rule, this involves a lot of trial and error and is often more an art than a science. "It takes away the art. It automates it, "said Joe Spisak, product manager of Facebook AI for PyTorch. "And our goal is to consume the latest research."
BoTorch is not the first Bayesian optimization tool. However, Facebook maintains that these existing libraries are difficult to expand and customize – and that they do not meet Facebook's needs.
Ax then exploits these features and provides all the management functions that allow BoTorch to find the optimal configuration for these models and allows developers to focus on the preparation of their services. On Facebook, Ax interfaces with A / B simulation and test tools, for example, and as these tools aim to automatically optimize the system, user participation is minimal. By performing the experiments, Ax can automatically choose the best optimization strategy. This can be Bayesian optimization, bandit optimization (another clbadic optimization strategy) or some other algorithm. Indeed, it should be noted that Ax is agnostic vis-à-vis the framework. While using BoTorch, researchers can also plug in their own code using the PyTorch and NumPy service interfaces.
Open sourcing tools on Facebook have become a norm at this time. PyTorch is a good example. Spisak also noted that for BoTorch, the team had decided to open the tool because it wanted to collaborate with the best researchers in the field (for example, by collaborating with Cornell University for this first version ). "How do we collaborate? How can we build an open community around what we do? You can not do this in a closed source. It must be open source, "he said.
Source link
