

The Google Brain team wants to help data scientists get their deep learning projects off the ground more easily with the release into open source today of a new software code library called Tensor2Tensor.
It’s designed to help researchers replicate the results of previous deep learning research projects while pushing the boundaries with new combinations of datasets, models and other factors. Deep learning, a branch of machine learning that uses some aspects of how the brain works to allow computers to learn on their own, is responsible for recent breakthroughs in speech and image recognition for applications such as language translation and self-driving cars.
The rapid pace of new developments and the huge number of variables in AI-based research makes it almost impossible for experiments run with distinct settings so it’s hard to run new experiments with the same models and compare the results to previous research. According to Google, that’s holding back the progress of AI research.
Tensor2Tensor is designed to make it easier for data scientists and researchers to maintain best practices while carrying out their research. Google describes T2T as an open-source system for training deep learning models in TensorFlow, which is the open-source deep learning framework it created.
“T2T facilitates the creation of state-of-the art models for a wide variety of ML applications, such as translation, parsing, image captioning and more, enabling the exploration of various ideas much faster than previously possible,” wrote Łukasz Kaiser, senior research scientist on the Google Brain team.
The library can do that because it comes equipped with an assortment of useful ingredients for deep learning experts such as hyperparameters, datasets, model architectures and learning rate decay schemes. These components can all be swapped in or out in a modular fashion without wrecking anything. This means researchers can add new models and datasets at any time without causing any disruption to their projects.
T2T also comes packaged with datasets used in some of Google’s own recent research projects. These include “Attention Is All You Need” and “One Model to Learn Them All.” The T2T library and datasets have been made available on GitHub, so budding data scientists can immediately get cracking on training their own deep-learning applications.
THANK YOU