Building Better Neural Models.

Collaboration between Amazon and Microsoft. Is this part of their broader collaboration for assistants we heard about recently?

Gluons for Open AI

AWS, Microsoft Team on ‘Open AI’     by George Leopold
Cloud services rivals Amazon Web Services and Microsoft are collaborating on a deep learning library designed to accelerate development of models used to construct neural networks.

The open source project dubbed "Gluon" would deliver an API that leverages a collection of neural network components to define machine-learning models. The partners said Thursday (Oct. 12) the Gluon interface will look more like "traditional code" since models can be tweaked like other data sets.

While the interface primarily targets developers new to machine learning, the partners added that Gluon would also allow data veterans to leverage dynamic neural network graphs to build and train new models without sacrificing speed, Matt Wood, AWS general manager for AI, noted in a blog post. The interface can be used to develop models running on the cloud, edge devices and mobile apps.

Further, they said Gluon code could be used to build deep learning components such as convolutional networks for applications ranging from object detection and speech recognition.

AI development, "especially deep learning models, isn’t easy," Microsoft's Eric Boyd added in a separate post. "It can be a fairly daunting and specialized practice for most data professionals. We believe bringing AI advances to all developers, on any platform, using any language, with an open AI ecosystem, will help ensure AI is more accessible and valuable to all."

The deep learning library is designed to help define, debug and maintain deep neural networks for cloud, edge and mobile applications. Gluon (the name refers to the elementary particle that serves as a link for the so-called "strong force" in quarks) includes a library designed to simplify development tasks such as defining complex structures underlying deep learning models. The approach allows users to reuse existing building blocks in the library.

Neural networks based on Gluon can then use AWS NASDAQ: AMZN) and Microsoft (NASDAQ: MSFT) distributed training tools the partners said could be scaled to more than 500 GPUs to significantly reduce training time. Optimized inference would also allow deep learning models to run on lower end hardware, they added. .... "