As enterprises increasingly pilot AI technologies, tiny machine learning, or TinyML, is emerging as the preeminent way to cut down on the resources required for deployment. TinyML is a machine learning technique that can be implemented in low-energy systems, like sensors, to perform automated tasks. The technology is still very clearly AI, but with lower power usage, costs, and often without the need for an internet connection.

Applications for TinyML run the gamut, but the most popular range from factories, retail, and agriculture. In the manufacturing sector, TinyML can prevent downtime by alerting workers to perform preventative maintenance based on equipment conditions. And in farming, TinyML can monitor the vitals of livestock to help identify the early signs of disease.

A number of startups offer products designed to help enterprises implement TinyML solutions, but among the most visible is Edge Impulse. Launched in 2019, Edge Impulse provides a platform and services for developing devices that leverage embedded AI and machine learning. It claims that nearly 30,000 developers from thousands of companies including Oura, Polycom, and NASA have created upwards of 50,000 custom projects using Edge Impulse solutions, building industrial, logistics, consumer, and health solutions.

Advancing TinyML

Edge Impulse was founded two years ago by Jan Jongboom and Zach Shelby. Jongboom previously contributed code to Mozilla’s now-discontinued operating system, Firefox OS, and lead developer evangelism for several of Arm’s internet of things (IoT) platforms. Shelby comes from an investment background, having served as a member of the boards of Petasense and proptech startup CubiCasa.

Edge Impulse allows developers to collect or upload training data from devices, label the data, train a model, and deploy and monitor the model in a production environment. The platform supports development for machine learning for sensors, audio, and computer vision, specializing in TinyML industrial applications including predictive maintenance, asset tracking, and monitoring, and sensing.

“Accuracy — which is normally used to assess the performance of a machine learning model — only tells a very small part of the story. You need to know the strengths and weaknesses of your model, know when it misses events or when it triggers false positives,” Shelby told VentureBeat via email. “[That said,] machine learning has huge value potential for all businesses working with sensor related data, from saving cost and better service customers to enabling whole new generations of feature value.”

Above: Edge Impulse’s development dashboard.

Image Credit: Edge Impulse

To increase the efficiency of models trained on its platform, Edge Impulse uses a compiler that compiles models to C++. The company claims that this can reduce RAM usage by 25% to 55% and storage usage by up to 35% compared with rival approaches.

“[We’ve seen] enterprise applications including human key word detection on battery-powered devices in wearables, predictive maintenance in the smart grid, gesture recognition using radar in devices, monitoring critical refrigeration equipment in the field, field detection of eye diseases, monitoring of welding quality using audio, [and] construction and manufacturing safety monitoring using computer vision and sensors,” Shelby said. “We saw customers slow down at the beginning of the pandemic, as they were not on the sites where they needed to collect data, but business has picked up very strongly.”

Growth year

According to Gartner, by 2027, machine learning in the form of deep learning will be included in over 65% of edge use cases — up from less than 10% in 2021. Meanwhile, ABI Research predicts that the TinyML market will grow from 15.2 million device shipments in 2020 to 2.5 billion in 2030.

Reflecting the broader segment’s growth, Edge Impulse reports that the developer base on its platform increased by four times last year, with annual recurring revenue growing by three times. In related news, the company today announced that it raised $34 million in a series B round led by Coatue with participation from Canaan Partners, Acrew Capital, Fika Ventures, Momenta Ventures, and Knollwood Investment Advisory, tripling its valuation to $234 million and bringing its total capital raised to over $54 million.

Shelby says that the new funds will be used to expand Edge Impulse’s roughly-40-person team and its hardware partner network, which already includes Nvidia, Texas Instruments, Syntiant, and Synaptics. “We’ll use the money to accelerate even faster, significantly growing our developer ecosystem to 100,000 developers by the end of 2022, rapidly grow our solution engineering team to help customers reach success faster, expand our hardware ecosystem, and invest in new R&D making machine learning for sensor, audio, and computer vision more efficient,” he added. 

Edge Impulse competes with startups like CoCoPie, Neural Magic, NeuReality, Deci, CoCoPie, and DeepCube, among others. FogHorn is one of its closest direct competitors, delivering a range of edge intelligence software for industrial and commercial applications. Incumbents like Microsoft, Amazon, and Google also offer services — for example, Amazon Web Services’ IoT Greengrass — targeting edge AI development through their respective cloud platforms.