Xnor brings self service to AI development
May 29, 2019
Seattle-based Xnor.ai has launched a self-serve platform that lets developers, device creators and companies build smart, edge-based products without training or background in AI.
Called AI2Go, it contains more than a hundred fully trained models optimised to run on resource-constrained devices such as mobile devices, wearables, smart cameras and remote sensors. Its models are being used to build technology for retail analytics, smart homes and industrial IoT.
Before AI2Go, AI relied almost exclusively on expensive hardware running in the cloud and was restricted to a handful of companies. Even with the tools that were available, building AI products required knowledge and expertise in deep learning to design, train and implement. Deploying these models at the edge required looking at a whole host of constraints, including memory, power and latency, which made development for on-device AI almost impossible.
AI2Go could change the scale and speed at which AI products can be built. By offering hundreds of trained edge AI models with accuracy, it means developers no longer need to worry about data collection, annotation, training, model architecture or performance optimisation. They simply download the complete product and are ready to go.
“Xnor’s drag-and-drop approach to AI application design removes much of the pain for developers living at the intersection of hardware and software,” said Adam Benzion, co-founder and CEO of Hackster.io.
Enterprise users of Xnor will continue to benefit from its custom trained and performance models. In the coming months, the AI2Go platform will provide enterprises access to optimised models along with additional custom features including automated training and re-training, and performance optimisation for large-scale development teams.”
“Xnor AI2Go is the gold standard for optimising IoT machine learning,” said Gant Man, CIO of Infinite Red.
The release of AI2Go is a continuation of Xnor’s mission to bring AI everywhere to everyone. In 2017, Xnor demonstrated it could remove the cost barrier by running deep learning on $5 hardware. In 2019, it removed the barrier of power with solar powered AI. Now, with AI2Go, Xnor says it is removing the barrier of AI expertise.
“By providing access to deep learning that can readily run on-device, we believe we afford all companies, regardless of team, budget or hardware, the opportunity to participate in this new era of AI innovation,” said Ali Farhadi, co-founder and CXO of Xnor. “AI2Go enables this vision through a platform of a large number of models running on many devices that are able to operate under numerous constraints.”
Using AI2Go is said to be simple. First the user selects their preferred hardware – Raspberry Pi, Linux, Ambarella, Toradex and so on – then chooses an AI use case, for example a pet classifier for a home security camera, a person detector for a dashcam or a person segmenter for video conferencing applications.
Because AI2Go models are designed to run in resource-constrained environments, Xnor provides the user with the opportunity to tune their model for latency (milliseconds) and memory footprint (megabytes) to fit within the user’s set of constraints. Once the user has specified their constraints, the available models are listed, ranked by accuracy. The user can then download a Xnor Bundle, a module containing a deep learning model, and an inference engine.
Xnor also provides an accompanying SDK that includes access to code samples, demo applications, benchmarking tools and technical documentation that makes it simple to start building a smart application.
Founded in 2017, Xnor is headquartered in Seattle, Washington.