Google AdaNet relies on TensorFlow and AutoML to provide developers with the benefits of Ensemble Learning.
The Google AI team has released AdaNet, a TensorFlow-based framework for automatically learning machine learning models. The tool also builds on the work on the service AutoML, which should also allow developers without great expertise to enter the world of machine learning (ML). According to Google, AdaNet should not only be a framework for learning neural network architecture, but also be able to use Ensemble Learning to create better models.
In Ensemble Learning, frameworks combine different predictions of ML models for better results. In practice, however, the technology is hardly ever in use, since the training times are extremely long and the selection of the models requires great knowledge. However, Google argues that with increasing computer speed and specialized deep-learning hardware, the concept will become more prevalent. This is where AdaNet comes in.
In the blog post, Google continues to say that AdaNet ML users should save time by combining different networks of different depths and breadths. AdaNet uses an adaptive algorithm to learn the neural network architectures, add subnetworks and create a so-called ensemble. This should guarantee learning success.
In addition, AdaNet implements the TensorFlow Estimator interface and the TensorBoard, the first encapsulates training, evaluation and prediction to export them for machine learning models. The second is a visualization tool that allows you to monitor models and provide different metrics.
If you would like to try AdaNet, you will find the project on GitHub. The makers also provide a series of tutorials.