"A New Technique Enables AI Models to Continually Learn From New Data on Intelligent Edge Devices Like Smartphones and Sensors, Reducing Energy Costs and Privacy Risks"
Microcontrollers, which are small computers that can execute simple commands, are the foundation for billions of connected devices, ranging from Internet of Things (IoT) devices to automotive sensors. However, because low-power microcontrollers have significantly limited memory and no operating system, training Artificial Intelligence (AI) models on "edge devices" that operate independently of central computing resources is difficult. When a Machine Learning (ML) model is trained on an intelligent edge device, it can adapt to new data and make better predictions. For example, training a model on a smart keyboard could allow the keyboard to continuously learn from the user's writing but since training requires so much memory, it is typically performed on powerful computers in a data center before the model is deployed on a device. Because user data must be sent to a central server, this is more expensive and raises privacy concerns. To address this issue, researchers at MIT and the MIT-IBM Watson AI Lab created a new method for on-device training that requires less than a quarter of a megabyte of memory. Other connected device training solutions can use more than 500 megabytes of memory, far exceeding the 256-kilobyte capacity of most microcontrollers. The researchers' intelligent algorithms and framework reduce the amount of computation required to train a model, making the process faster and more memory efficient. In a matter of minutes, they can train an ML model on a microcontroller. This technique also protects privacy by keeping data on the device, which may be especially useful when the data is sensitive, as in medical applications. This article continues to discuss the researchers' new method that allows AI models to continuously learn from new data on intelligent edge devices while lowering energy costs and privacy risks.