Google LLC today announced a new tool called TensorFlow Lite Model Maker, which uses a technique known as transfer learning to adapt machine learning models to custom data sets.

TensorFlow Lite is a slimmed-down version of Google’s TensorFlow framework for training machine learning models. It’s a set of tools used by developers to run TensorFlow models on mobile, embedded and IoT devices. It enables the operation of on-device machine learning with low latency and a small code size.

Model Maker is said to combine machine learning concepts with an application programming interface that developers can use to train models in the TensorFlow Lite framework using just a few lines of code. It can then be used to deploy those models within on-device applications, Google said.

The tool is compatible with many of the models in TensorFlow Hub, which is a library of reusable machine learning modules. What Model Maker does is take those models, which have been trained for a specific task, and retrains them for another related task. It can do so with varying levels of accuracy, determined by whatever parameters the developer sets at the outset.

Developers can improve the accuracy of the new models by changing just a single line of code to adapt Model Maker’s architecture, Google said. Once the input data for a specific on-device AI has been loaded, Model Maker evaluates that model and then exports it as a TensorFlow Lite model.

Google said that it’s planning to enhance Model Maker to support even more AI tasks, including natural language processing and object detection. For example, it’s planning to add BERT, which is a training technique for NLP that enables question and answer applications.

The launch of Model Maker is just the latest in Google’s long-running effort to democratize artificial intelligence, said Constellation Research Inc. analyst Holger Mueller.

“Experience with AI and machine learning is not abundant, and data scientists are often too busy or too expensive to hire, so it always comes down to regular developers to operate and tune the models enterprises require,” Mueller said. “This is what Google is aiming at with Tensorflow Lite Model Maker, trying to get more people to use its technology to build their next-generation applications.”

Image: Google