The performance of non-invasive brain devices can be improved by combining them with AI.Credit: Jean-Pierre Clatot/AFP via Getty
A man with partial paralysis was able operate a robotic arm when he used a non-invasive brain device partially controlled by artificial intelligence (AI), a study reports1. The AI-enabled device also allowed the man to perform screen-based tasks four times better than when he used the device on its own.
Brain–computer interfaces (BCIs) capture electrical signals from the brain, then analyse them to determine what the person wants to do and translate the signals into commands. Some BCIs are surgically implanted and record signals directly from the brain, which typically makes them more accurate than non-invasive devices that are attached to the scalp.
Jonathan Kao, who studies AI and brain–computer interfaces at the University of California, Los Angeles, and his colleagues wanted to improve the performance of non-invasive BCIs. The results of their work are published in Nature Machine Intelligence this week.
First, the team tested their BCI by tasking four people — one with paralysis and three without — with moving a computer cursor to a particular spot on a screen. All four were able to complete the task the majority of the time.
When the authors added an AI co-pilot to the device, the participants completed the task more quickly and had a higher success rate. The device with the co-pilot doesn’t need to decode as much brain activity because the AI can infer what the user wants to do, says Kao. “These co-pilots are essentially collaborating with the BCI user and trying to infer the goals that the BCI user is wishing to achieve, and then helps to complete those actions,” he adds.
The researchers also trained an AI co-pilot to control a robotic arm. The participants were required to use the robotic arm to pick up coloured blocks and move them to marked spots on a table. The person with paralysis could not complete the task using the conventional, non-invasive BCI, but was successful 93% of the time using the BCI with an AI co-pilot. Those without paralysis also completed the task more quickly when using the co-pilot.
