Google today introduced AdaNet, an open source tool for combining machine learning algorithms to achieve better predictive insights. AdaNet is available today on the Tensor GitHub repository.
“AdaNet builds on our recent reinforcement learning and evolutionary-based AutoML efforts to be fast and flexible while providing learning guarantees,” Google AI software engineer Charles Weill said in a blog post. “Importantly, AdaNet provides a general framework for not only learning a neural network architecture, but also for learning to ensemble to obtain even better models.”
AdaNet uses an approach called ensemble learning to combine and improve algorithms, a method that previously required domain expertise or too much time for training, Weill said.
To make it easier to implement AdaNet, the framework plugs into the TensorFlow Estimator to bring essential information into a single place, as well as TensorBoard, which delivers visual feedback when an AI model is being trained.
AdaNet ensures learning guarantees for the ensemble models it creates by learning the architecture of neural networks, then adding subnetworks to them.
Machine learning practitioners who want more control of the process can use TensorFlow APIs to define their own subnetworks, customize loss functions, or toggle other settings.
Additional details about how AdaNet works can be seen in this published paper presented last year at the International Conference on Machine Learning.
The release of AdaNet today is the latest step forward in AutoML, Google’s automated way to train and deploy neural networks. Google Cloud Platform introduced AutoML for translation, computer vision, and natural language processing this summer, as well as Cloud AutoML for building custom AI models in January.