Machine Learning in Edge Devices: French Start-up Takes IoT to the Edge

Smart Blog

Machine Learning in Edge Devices: French Start-up Takes IoT to the Edge

Cartesiam, a small company focused on machine learning in edge devices, has won the IoT World 2020 Startup Elevate Pitch-Off award at the 2020 IoT World conference. The competition required startups to pitch their products to a panel of industry leaders, investors and media.

Founded in 2016 in France, Cartesiam concluded that projections for the Internet of Things (IoT) market at that time were misguided.

Everyone was talking about IoT — deploying tens of billions if not hundreds of billions of devices,

recalled Marc Dupaquier, co-founder and managing director of the company. But despite the word “smart” to describe IoT-enabled devices and environments, many rely on the cloud to perform routine workflows. Yet using the cloud to process the bulk of IoT-based processing tasks will “not be economically viable,” Dupaquier concluded.
In Dupaquier’s view, IoT devices should have sufficient resources to understand operational parameters natively. A cloud outage shouldn’t prevent an IoT-enabled industrial machine, for instance, from sending an alert warning of a pending problem.
While such commonsense functionality has long been a promise of IoT technology, few companies have mature AI deployments. Fewer than one in 20 organizations have extensively integrated AI in offerings or processes, according to a survey from MIT Sloan Management Review.
The concept of performing artificial intelligence (AI) at the edge has grown in popularity recently. Adoption, however, remains at an early phase — especially when it comes to doing edge computing on resource-constrained hardware.

Machine Learning in Edge Devices Team Cartesiam

In the middle of 2017, Cartesiam began developing a low-cost local computing platform that performs training and inference tasks at the edge. For hardware, the startup decided to focus on microcontroller units (MCUs), which range in cost from a few cents to a few dollars per board. The relatively limited horsepower of MCUs required bespoke machine learning and signal processing algorithms and logic; the bulk of recent AI research has relied on cloud architectures.
The company now has some 500 million algorithms for running AI operations on MCUs. While the company wrote algorithms manually at first, it was eventually able to generate scores of permutations of algorithms for narrow applications.
Predictive maintenance at the edge in industrial settings was an early focus area for Cartesiam. More recently, that focus has broadened to predictive maintenance for home appliances.
The ability to make local decisions is important for many applications requiring immediate action, but that doesn’t mean edge architecture is warranted in every case. A machine in a factory with access to unlimited power, connected to a nearby data center or cloud with high bandwidth may not need to have as much processing at the edge.
As for applications, Cartesiam is looking to expand its business to include wearables as well as industrial machines and home appliances. The analytics capabilities of many mainstream wearable devices are limited in Dupaquier’s estimation, who is a runner who has tested with multiple wearables.

The first time I played tennis, I wore an expensive wearable from a well-known global brand. Afterward, I received a text message from the wearable device that said something like: ‘Congratulations on a great swim,’

Dupaquier said.

My friends and colleagues told me I am probably not a very good tennis player if this algorithm confuses playing tennis with swimming.

Author: Tim Cole
Image Credit: Cartesiam

Leave a Reply

Your email address will not be published. Required fields are marked *