Philosophy

A2ML is an source project to Automate Automated Machine Learning. It was started by the team of Auger.AI. But it is not an attempt to make Auger.AI's own API a standard. In general we have erred towards using the option names from other AutoML providers. In particular we found Microsoft Azure AutoML's vocabulary to be reasonable and descriptive. So we adopted much of their vocabulary in A2ML.

Why Are We Doing This

The team behind Auger.AI has a history of building free useful services and figuring out how to monetize them later. We felt that the AutoML industry desperately needed a common API. The dominant players in AutoML have particularly poor performance in finding optimal algorithms and hyperparameters, due to their unsophisticated grid search approach. AutoML needs innovation and disruption. An open, common API makes this much easier to happen.

We are also offering a paid hosted service that manages the connections to AutoML services from various vendors. Just like you pay for hosted instances of other technologies (Spark from DataBricks, Docker, Redis) we offer a hosted A2ML service that manages your connections and instances of AutoML providers from various vendors.

Review and Monitoring

A2ML also introduces the concept of a Review phase to help users monitor that their models remain accurate. Auger.AI includes charts and diagnostics of a model's ongoing accuracy. While the Auger Review service is paid, the base code for aggregating accuracy (comparing predictions and actual results and several measures of ongoing data values) is in fact free as part of A2ML. This should allow accuracy to be provided for AutoML services which don't provide this capability (today that is most other other AutoML providers). We believe this will result in more accurate AutoML models in general over the lifetime of their usage.