Researchers at the University of Technology Sydney (UTS) say the new DeWave technology is noninvasive. Until now, efforts to read brain signals required implanted electrodes or costly MRI machines.
Now, to record thoughts and turn them into text, all you need is a special cap that fits snugly on the head. The cap’s sensors capture brain waves to produce electroencephalograms (EEG), which the system decodes using the DeWave codebook.
What the researchers found
Researchers tested the system on two dozen volunteers. The volunteers silently read text while wearing the caps, which recorded their brain waves. The recorded EEG data were decoded into text by the AI.
The team says that with further improvements, DeWave could help stroke and paralysis patients and assist in controlling bionic hands or robots.
Computer scientist Chin-Teng Lin called the research an innovative attempt to translate raw EEG signals directly into language. This marks a significant breakthrough in the field.
Although in experiments conducted by Lin and his colleagues DeWave reached only about 40 percent accuracy, that’s still three percentage points higher than previous invasive models. The team’s near-term goal is to push the AI’s translation accuracy to at least 90 percent.
The system is best at verbs
After training, the DeWave encoder converts EEG signals into a discrete code that is matched to words in the DeWave codebook based on similarity.
“This is the first model that has incorporated discrete coding methods into the process of translating brain waves into text, representing an innovative approach to neural decoding,” explained Chin-Teng Lin. In his view, integration with large language models “also opens new horizons in neurobiology and artificial intelligence.”
The team used language models BERT and GPT, testing them on datasets from eye-tracking and brain-activity recordings during reading. As a result, the system learned to match brain wave patterns to words in the codebook, the publication reported. Science Alert.
The best results came when the artificial intelligence system translated verbs. However, nouns were the hardest for the system. For example, the words “person” and “author” were treated by the AI as nearly identical concepts.
Lead author Yicun Duan suggests this may happen because the brain produces similar neural patterns when processing semantically similar words. “In any case, the model shows promising results, aligning keywords and forming similar sentence structures,” he added.
Researchers are focused on future studies of the system. “Translating thoughts directly from the brain is a valuable but complex task that requires significant ongoing effort,” the scientists wrote in their report.
