Using ONNX to Complete Machine Learning Projects Faster with Less Trial-and-Error

Posted by Jason Behrmann, PhD on Nov 12, 2020 10:00:00 AM

Given that commercial applications of machine learning have only recently entered the limelight, methods for developing the technology are still, for the most part, based on ad-hoc solutions. Common standards for commercial applications of machine learning remain far and few, which has many businesses wasting time and resources developing and maintaining homegrown hardware and software for their AI solutions. This not only increases the costs of innovation, but the multiplicity of technology formats also makes it impossible for AI systems and technology tools to work together.

Well aware of the problem, leaders in the AI sector are calling for the implementation of industrial standards in machine learning and the tools used to advance AI innovation. One initiative of particular import is by the ONNX open-standard community from the pioneering organization for open-source technology tools in AI, The Linux Foundation AI. Short for Open Neural Network eXchange, ONNX provides a standard format to represent and share machine learning models. This standard makes diverse software and hardware compatible. In turn, this enables professionals from around the globe to implement diverse technology tools on a given project when developing innovations in AI.

Since 2019, a growing list of prominent tech companies are employing ONNX in popular AI products and services. This list now includes us at Zetane Systems, where we implement ONNX within our software for industrial applications of machine learning, the Zetane Engine.

To initiate the introduction of Zetane to its community, members of the ONNX steering committee invited Zetane to present with fellow partners at the upcoming ONNX Community LF AI Day this October. In the presentation below, the CTO of Zetane, Patrick St-Amant, describes how the Zetane Engine leverages ONNX to shorten development cycles for machine learning projects. The solution resides in new abilities to change abstract, “black-box algorithms” into detailed visual displays of the inner workings of artificial neural networks. The second focus of the talk will explain how the open standard provides more transparent means to optimize the performance of artificial neural networks, which in turn reduces the need to conduct time-consuming trial-and-error strategies.

If text is more your thing, you can read a transcript of the presentation below.