EML and Tiny machine learning (TinyML) are similar in that they both involve deploying ML models on small, resource-constrained devices, such as smartphones or IoT devices, rather than on powerful servers or in the cloud. However, the terms may be used slightly differently depending on the context. Tiny ML generally refers to the use of ML on very small and low-power devices, such as microcontrollers, while EML refers to the deployment of ML models on devices located at the “edge” of a network, closer to the source of data, such as microcontrollers.
Industrial and manufacturing, smart cities and consumer applications are driving the demand for EML. Bringing learning algorithms to millions of small, low-powered devices at the edge opens up endless possibilities for intelligence and valuable intelligence. Each EML device is context aware and understands its environment, meaning that individualized and optimized approaches can be carried out regarding predictive maintenance or personalized healthcare initiatives, for example.
On-device analytics allows edge devices to process data without having to send it to the cloud, which can result in sizeable network latency. This makes EML a great fit to support real-time environments, such as crop monitoring and production lines.
Keeping data at the edge also reduces the risk of data breaches as data doesn’t need to be transferred and stored in cloud servers. This is important in areas where sensitive data is being used, such as healthcare. In addition, EML also has a much lower carbon footprint than what cloud AI offers, making it score in terms of sustainability.
“ML looks for commonality in large amounts of data; this requires expensive processing power. Edge machine learning turns this upside down. It starts at the data source and can learn from the processes from the data it sees,” explains Anders Alneng, Co-Founder and Vice President of Sales at Ekkono Solutions, a pioneer in edge ML software.
Putting AI on the edge
ABI Research forecasts that 2.5 billion devices will ship with TinyML in 2030, driven by the growing focus on low latency, advanced automation and the availability of low-cost, ultra-power-efficient AI chipsets.
“Since AI is deployed to make immediate critical decisions such as quality inspection, surveillance and alarm management, any latency within the system may result in machine stoppage or slowdown causing heavy damages or loss in productivity. Moving AI to the edge mitigates potential vulnerability and risks such as unreliable connectivity and delayed responses,” explains Lian Jye Su, Principal Analyst at ABI Research.
“Everything is getting connected. But the most important thing is that this is the first time OEMs can maintain relationships with their products when they leave the factory. This requires automation, and IoT automaton equals EML,” adds Ekkono’s Alneng. “Now you can make machines better when they leave the factory and learn from their production environment. It supports maintenance and optimization, but the effect is also huge regarding sustainability. There is less wear and tear on the product, and it can be more energy efficient – and you can do individualized per machine leading to mass customization. This was not possible before.”
EML is not an answer to every problem, however. EML is not the ideal approach, for example, where you need a lot of memory to process a large amount of data, such as in natural language or refined image processing. “Here, traditional machine learning in the cloud is a better solution,” explains Kasra Mohaghegh, Data Engineer at Orange Business, “But EML has its place in many cases that require low-latency, real-time interactions.”
EML in action
Ekkono recently utilized EML to help a manufacturing company adopt continuous efficiency changes for a customer’s heat exchangers.
The plates in the heat exchangers get fouled and need maintenance, varying from every couple of weeks to years, depending on the liquids used. The company was finding it impossible to plan maintenance schedules without having accurate insight into fouling models.
Ekkono used virtual sensors to measure data from physical sensors and create a component fouling estimation. This has allowed continuous condition monitoring, preventing unplanned production stops and scheduling planned maintenance for the manufacturer.
Standard batch method training requires a large amount of historical data, and often there isn’t enough failure data to support machine learning models. By using incremental learning, this solution does not need historical data to learn from. Instead, it leverages real-time data for accurate insight. Any deviation from the normal state indicates that something has changed within the device.
The next step is federated learning
Federated learning (FL) is a part of machine learning that is gaining much attention. It allows shared AI models to be trained collaboratively on decentralized data at the edge without sharing or putting it in a central location. The model is improved iteratively.
“This approach allows for an overall general model to be created from all the characteristics of the small models,” explains Mohaghegh. Potential applications include autonomous vehicles and wearable devices. Federated learning is also of specific interest to government organizations, FinTechs and multinationals who want to retain data ownership.
“As digital transformations continue to drive AI/ML initiatives, companies will need to train robust and performant models across multiple locations without moving the data,” explains Ritu Jyoti, Group Vice President of AI and Automation Research Practice at IDC. “FL is expected to make significant strides forward and transform enterprise business outcomes responsibly.”
A whole new data source
The beauty of EML is that it is possible to develop machine-learning models on a host of embedded devices, from mobile embedded systems to tiny microcontrollers.
Training at the edge requires less data and less bandwidth. Each device changes from simply connected to smart, opening the potential to utilize new data sources and provide individualized learning. “It is like having a personal trainer and a doctor with you all the time, optimizing and understanding each machine’s health continuously,” says Ahneng.
To find out more about EML and its potential, listen to the Orange Business webinar: Making IoT devices smarter – the next frontier for machine learning on the edge.
Jan has been writing about technology for over 22 years for magazines and web sites, including ComputerActive, IQ magazine and Signum. She has been a business correspondent on ComputerWorld in Sydney and covered the channel for Ziff-Davis in New York.