2023-12-15 02:40:09
The Center for Human-Centered Artificial Intelligence, GrapheneX-UTS, at the University of Technology Sydney (UTS), has achieved a world first by developing an AI that can read thoughts and turn them into text.
According to the news, this device, portable and less invasive, can help people who cannot speak due to illness or injury. In particular, a stroke or paralysis. She also promotes smooth communication between humans and machinessuch as controlling an arm or a bionic robot.
UTS research into AI that reads minds into text applauded at NeurIPS 2023 conference
The study was honored as the lead paper at the NeurIPS conference, a annual event dedicated to cutting-edge research in artificial intelligence and machine learning. Professor Emeritus CT Lin led the research. But he collaborated with Yiqun Duan, the first author, and Jinzhou Zhou, a doctoral student from the Faculty of Engineering and Computer Science at UTS.
Concretely, the participants of the study read passages of text silently. However, they wore a cap recording brain electrical activity using an electroencephalogram (EEG). The EEG waves are then segmented into distinct units by an AI model called DeWave, created by the researchers. DeWave then deciphers the EEG signals into words and sentences by learning from a vast amount of EEG data.
A unique method of brain-text translation without surgery or MRI
“This research work represents a pioneering advance in directly converting raw EEG waves into language. Our mind-reading AI indeed marks a significant breakthrough in this field,” Lin emphasized.
” We are the first to integrate discrete coding techniques into the brain-to-text translation process. We have also introduced an innovative approach to neural decoding. Integration with large language models also opens up new perspectives in neuroscience and artificial intelligence,” he added.
Previous technologies require the translation of brain signals into language. The researchers faced surgery to implant electrodes in the brain. This is also the case with the Elon Musk’s Neuralink project. In other words, we are talking regarding MRI scanning. However, this approach is more cumbersome, expensive and difficult to use.
These methods also have difficulty converting brain signals into segments. Particularly at the word level, where they need additional help. Including eye tracking which limited their practical application. On the other hand, this device, powered by AI, can be used to read thoughts with or without eye-tracking.
An accuracy rate of less than 50%: is this enough to get ahead of humankind?
The UTS study involved the participation of 29 subjects. Which made it possible to reinforce its robustness and adaptability compared to previous decoding technologies.
By using EEG signals captured by a cap rather than electrodes implanted in the brain, the signal is certainly noisier. However, in terms of EEG translation, the study has demonstrated exceptional performancesurpassing previous benchmarks.
In a world-first, researchers from the GrapheneX-UTS Human-centric Artificial Intelligence Centre at the University of Technology Sydney (UTS) have developed a portable, non-invasive system that can decode silent thoughts and turn them into text.https://t.co/lP0zj9Q7Qn
— Sandbird (@Sandbird_) December 13, 2023
“Our model excels more at associating verbs than nouns. Concerning the names, we have observed a preference for pairs of synonyms rather than exact translationslike the man instead of the perpetrator, Duan explained.
“This is likely because, when processing these words, semantically similar terms can generate comparable brain wave patterns. Despite these challenges, our model produces meaningful results by aligning keywords and forming similar sentence structures “, he added.
The current translation accuracy rate is around 40% on BLEU-1. The BLEU score is a measure of similarity between machine-translated text and a set of high-quality reference translations, ranging from zero to one. Researchers aspire to improve this figure to bring it closer to traditional translation or voice recognition programswhich generally reach almost 90%.
1702613369
#reads #minds #wary
