AI at the Edge: New Machine Learning Engine Deploys Directly on Sensors

August 03, 2020 by Maya Jeyendran

ONE Tech's upgraded MicroAI Atom brings AI training and operation down to a single device.

ONE Tech, an AI and ML-driven company specializing in Internet of Things (IoT) solutions for network operators, enterprises, and more, has announced new capabilities of its MicroAI Atom product.

MicroAI Atom is part of ONE Tech’s Micro AI product line, and it now has the ability to train and run AI models at the edge, enabling a variety of individuals and entities to “reduce the costs of bringing intelligence to the edge and endpoint by at least 80 percent.” 


What Is MicroAI?

According to ONE Tech’s press release, MicroAI is a machine learning algorithm that is embedded into microcontroller units and operates a recursive analysis of device behavior. More specifically, MicroAI collects data from internal device sensors and utilizes a semi-supervised learning approach to come up with a complete view of device behavior.

Semi-supervised learning refers to a machine learning method that makes use of some labeled data with a much larger amount of unlabeled data during the model’s training phase. Other machine learning methods can either employ a fully supervised approach that would consist of using only labeled training data or a fully unsupervised approach that would consist of using only unlabeled training data.


MicroAI Atom interacts with device sensors to obtain data it uses to inform its AI models.

MicroAI Atom interacts with device sensors to obtain data to inform its AI models. Image used courtesy of ONE Tech

MicroAI Atom, unlike existing edge-based microcontrollers used for device management, is employed directly in IoT devices and sensors, thereby enabling real-time alerts, all while increasing the device’s security profile.

This product compares well to Cartesiam's NanoEdge AI Studio, which also allows designers to integrate machine learning directly into edge devices that are based on Arm Cortex-M MCUs.


Security Benefits of Edge AI Training

AI models have largely been trained by leveraging cloud-hosted, GPU-based servers; while cloud computing has numerous, well-documented advantages, security concerns regarding the consequences of cloud-hosted data and models being intercepted certainly exist. 

Micro AI Atom’s training AI models directly at the network edge will allow users of this technology to bypass these security concerns altogether since data no longer needs to be transmitted to the cloud. 


Different Models for Different Devices

Security benefits aren’t the only benefits to emerge from the ability to train AI models directly at the edge. AI models trained in the cloud are typically “pushed down to the edge,” implying that each existing device will be subject to modifications made by the same AI model.


Typical MicroAI ATOM deployment process

Typical MicroAI ATOM deployment process. Image used courtesy of ONE Tech

Even if certain devices have much in common, it’s important to note that they need not have similar performance requirements. For example, two otherwise identical devices might be subject to use in dramatically different weather conditions and may, therefore, require totally different modifications for each perform optimally.

Thus, having a different model for each device can lead to significant performance improvements.


Selectiveness in Transmitting Data

The ability to train directly at the edge also has advantages when it comes to creating and transmitting large amounts of data. 

Millions of data points captured from the device no longer need to be migrated to the cloud or another location for analysis, and furthermore, MicroAI Atom only outputs data processed by its algorithm, thereby reducing the amount of data by several orders of magnitude. 



Do you work with machine learning models directly deployed on a device? What are the benefits and limitations of this integrated design method? Share your experience in the comments below.