Breakthrough Could Lead to Development of Energy-Efficient AI

Running artificial intelligence applications on local devices comes with tons of advantages.

For one, you won’t require a network connection with to run voice assistant software such as Alexa on devices. Also, manufacturers can create privacy-friendly electronics that can store and process data locally.

There’s just one problem.

Current AI applications are not power-efficient enough to process data locally on smart devices. As a result, applications such as speech recognition and gesture recognition rely on cloud connection to work.

But that could change soon, thanks to a recent study from Centrum Wiskunde & Informatica (CWI).

The researchers have made a mathematical breakthrough that can make AI application a thousand times more power efficient. What’s more, they’ve made the underlying mathematical algorithms open source.

The researchers described their work in their yet to be peer-reviewed paper published in arXiv.org.

A Mathematical Breakthrough for Energy-Efficient AI Applications

Spiking neural networks have been around for a while.

However, they can be challenging to handle from a mathematical perspective. This, in turn, makes it difficult to put such a neural network to practice.

However, such limitation didn’t stop the researchers. In the recent mathematical breakthrough, the team developed a learning algorithm for a spiking neural network.

The algorithm offers two significant advantages over current models.

The neurons in the network communicate less frequently
Individual neurons execute fewer calculations

Thanks to these factors, the team was able to develop a more energy-efficient AI application.

In a statement about the project, principal investigator, Sander Bohté explained:

“The combination of these two breakthroughs make AI algorithms a thousand times more energy efficient in comparison with standard neural networks, and a factor hundred more energy efficient than current state-of-the-art neural networks.”

The breakthrough could take AI applications to the next level. For example, it becomes possible to put a more elaborate artificial intelligence in chips to enable a more extensive application.

First, new types of chips are necessary to run spiking neural networks efficiently in the real-world. Luckily, various manufacturers are already working on creating such chips.

“All kinds of companies are working hard to make this happen, like our project partner IMEC/Holst Center,” said Bohté.