Neural network standards launched by Khronos Group to accelerate cross-platform deep learning

SAN FRANCISCO, CA – The Khronos Group has announced two standard initiatives that will address the development of neural network technology, the Neural Network Exchange Format (NNEF) and the convolutional neural network (CNN) topology extension to the OpenVX working group. The creation of these working groups is intended to accelerate the expansion of cross-platform deep learning tools, engines, and applications while reducing deployment challenges.

The NNEF working group is tasked with creating an API-independent open standard file format for the exchange of deep learning data between inferencing engines and network training systems. NNEF data interchange format is designed as a tool that can be used to create a trained network and run that network on other toolkits or inferencing engines. The NNEF standard comprises commonly used operations such as convolution, normalization, and pooling, as well as neural network structure, data formats, and formal network semantics to enable reliable import/export without enforcing how exported networks are trained or imported networks are executed.

The OpenVX Neural Network extension is an high-level architecture specification for executing CNN inferences in OpenVX graphs, and defines a multi-dimensional tensor object data structure that can be used to connect the various layers of a neural network. Network layers are represented as OpenVX nodes, with neural layer types including activation, convolution, fully connected, normalization, pooling, and soft-max, with nine different activation functions. The OpenVX Neural Network extension allows neural net inferences to be combined with vision processing operations in the same OpenVX graph, and an API has been defined in OpenVX that complements the Neural Network extension by allowing the import and export of OpenVX objects. The OpenVX Neural Network extension remains portable and processor-independent, and has been released in provisional form.

“Khronos’ efforts to standardize a universal CNN description exchange format will speed the availability of universal tools for converting trained CNNs to the inference domain,” said Dino Bekis, vice president of product marketing for the IP Group at Cadence. “The extensions to OpenVX graph descriptions will enable more seamless deployment of both imaging and vision algorithms in deeply embedded devices.”

Khronos expects that NNEF files will be able to represent all aspects of an OpenVX neural network graph, and that OpenVX will enable import of network topologies via NNEF files through the Import/Export extension once the NEFF format definition is complete. Industry participation is encouraged, and more information about the NNEF and OpenVX Neural Network extension working groups can be found at and, respectively.