AdaNet (program for consolidation of algorithms of machine learning)

2018: Announcement

At the end of October, 2018 Google submitted the software with open initial kodomadanet which integrates algorithms of machine learning for obtaining more effective results of predictive analytics.

AdaNet is based on our last achievements in training with a reinforcement signals from the environment of interaction and efforts on AutoML development to be fast and flexible and also to set skills of training — the software engineer of division Google AI Charles Weill whose words he is brought in the blog of the company says. — It is important to note that AdaNet provides the general platform not only for studying of architecture of neural network, but also for training of ensemble to receive even more best models.

Google released the tool for consolidation of algorithms of machine learning

Vayll noted that the new solution is used by a so-called training method of ensemble for consolidation and improvement of algorithms which to an exit of AdaNet required expert knowledge in this area and too much time for training.^[1]

For the simplified AdaNet implementation the platform is connected to the TensorFlow emulator for providing important information in one place and also to a set of web applications TensorBoard which provides a visual feedback when training AI model. AdaNet provides the guaranteed training in ensemble models which he creates by studying of architecture of neural networks and the subsequent adding of subnets to them. Thanks to AdaNet machine learning specialists who want to control this process can use TensorFlow API for determination of own subnets, creation of the loss functions (Loss function) or switching of other parameters.

Source codes of AdaNet are placed on the GitHub portal in Tensor repository.^[2] More detailed information on operation of the tool published in article provided at the International conference on machine learning in Stockholm    (International Conference on Machine Learning).^[3]

Notes