Khronos Group creates standard for data scientists
Wed 20 Dec 2017
A consortium of hardware and software companies has released a standard for ease of sharing among data scientists.
The Khronos Group has spent more than a year creating the Neural Network Exchange Format (NNEF) 1.0 Provisional Specification to ease the exchange of trained neural networks.
As it stands, machine learning deployment can become fragmented. The 1.0 specification released by Khronos aims to alleviate this by providing data scientists and engineers with a standardised way of exchanging neural networks training tools and inference engines.
Having been created by data scientists and engineers, these training tools and inference engines can, in turn, be used by applications across various devices and platforms.
NNEF does this by having all the elements of a trained neural network, such as the structure, operations and parameters, but acting independently of the training tools used to make it and the inference engine used to carry it out.
The Khronos group argues that a stable, flexible and extensible standard is extremely important for manufacturers, so that they are able to deploy neural networks onto edge devices on a large scale without major issues.
“The field of machine learning benefits from the vitality of the many groups working in the field, but suffers from a lack of common standards, especially as research moves closer to multiple deployed systems,” said Peter McGuinness, NNEF work group chair.
“Khronos anticipated this industry need and has been working for over a year on the NNEF universal standard for neural network exchange, which will act as the equivalent of a Pdf for neural networks.”
NNEF 1.0 has been designed so that it can be used across various tools and engines like Torch, Caffe and TensorFlow, as well as others. According to Khronos, future work on the platform will be built in a predictable way, allowing it to keep up with the fast-moving field of machine learning while remaining stable.
The standard has received support from a number of major industry names, including chip maker Qualcomm.
Jeff Gehlhaar, VP of technology at Qualcomm, said: “As a Khronos member, Qualcomm believes consolidation will help growth in this area, and is supportive of standards for representation of neural network models, such as Khronos Neural Network Exchange, which streamlines the migration from cloud to device.”
As a first release, those at Khronos are welcoming feedback on the organisation’s GitHub repository.