New AI models will soon learn on high-end devices

pte20221005003 research/development, technology/digitization

MIT scientists develop process that saves energy costs and improves data protection

Machine learning model on high-end machines: allows adaptation to new data (Image: mit.edu)

Cambridge (pte003/05.10.2022/06:10) –

Thanks to new technology, AI models will constantly learn from new data found on smart devices such as smartphones and sensors. This reduces energy costs and data protection risks. The technology is developed by researchers Massachusetts Institute of Technology (MIT) and des MIT-IBM Watson AI Lab. Details will be presented at the Neuroinformation Processing Systems Conference in a few weeks.

embarrassing workout

Microcontrollers that carry out simple commands are the basis for billions of connected devices, from the Internet of Things (IoT) to sensors in cars. However, cheap, low-power microcontrollers have very little memory and no operating system—a challenge when training artificial intelligence (AI) models on high-end hardware that operates independently of central computing resources.

When a machine learning model is trained on a smart edge device, it can adapt to new data and make better predictions. For example, if a model is trained on a smart keyboard, it can constantly learn from what the user is typing. However, the training process requires a large storage space which is usually performed on powerful computers in the data center before the model is deployed to the machine. This is expensive and raises privacy issues as user data must be sent to a central server.

Faster and more efficient

To solve the problem, the MIT team is developing a new technology that allows training on the device with less than a quarter of a megabyte of storage space. Other connected hardware training solutions require more than 500MB of memory, far exceeding the 256KB capacity of most microcontrollers. The researchers’ clever algorithms and framework reduce the computational effort required to train a model, making the process faster and more memory efficient. This technique allows training a machine learning model on a microcontroller in minutes.

According to experts, the technology also protects privacy, as the data remains on the device, which is especially useful for sensitive content such as medical applications. It can also allow a model to be customized to the user’s needs. In addition, the framework maintains or improves the accuracy of the model compared to other training methods. “Our study enables IoT devices not only to draw inferences, but also to constantly update AI models with newly collected data, paving the way for lifelong on-device learning. Using low resources makes deep learning more accessible and can have an impact,” says lead author Song. Han:

(End)