Meta and Researchers Unveil AI Models That Convert Brain Activity Into Text with Unmatched Accuracy
Serving tech enthusiasts for over 25 years, TechSpot offers tech analysis and advice you can trust. Recently, Meta, in collaboration with international researchers, announced major advancements in understanding human intelligence through groundbreaking studies. These impressive strides involve creating AI models capable of reading and interpreting brain signals to reconstruct typed sentences and understand how thoughts transform into spoken or written words.
The first study, conducted by Meta’s Fundamental Artificial Intelligence Research (FAIR) lab in Paris alongside the Basque Center on Cognition, Brain, and Language in San Sebastian, Spain, demonstrates a remarkable breakthrough. This research showcases the ability to decode sentence production from non-invasive brain recordings. Using cutting-edge techniques like magnetoencephalography (MEG) and electroencephalography (EEG), researchers recorded brain activity from 35 healthy participants as they typed sentences. The system employs a sophisticated three-part architecture: an image encoder, a brain encoder, and an image decoder. Initially, the image encoder builds a set of diverse representations independently from the brain. Subsequently, the brain encoder learns to align MEG signals with these image embeddings. Finally, the image decoder generates plausible images based on these brain representations.
The outcomes are impressive, with the AI model successfully decoding up to 80 percent of characters typed by participants whose brain activity was recorded with MEG. This efficiency is at least twice that of traditional EEG systems, presenting new horizons for non-invasive brain-computer interfaces. Such interfaces could potentially restore communication for individuals who have lost their ability to speak.
The second study delves into how the brain transforms thoughts into language. By employing AI to interpret MEG signals while participants typed, researchers identified precise moments when thoughts become words, syllables, and letters. This research reveals that the brain crafts a sequence of representations, transforming from the abstract meaning of a sentence to specific actions like keystrokes. Intriguingly, the study also demonstrates that the brain utilizes a ‘dynamic neural code’ to link successive representations while retaining each over extended durations.
Despite the promising outcomes, certain challenges persist before this technology can be clinically applied. The decoding process is not yet perfect, and MEG requires individuals to be in a specially shielded room and remain still. The MEG scanner is notably large, costly, and requires operation in a shielded space, as the Earth’s magnetic field is immensely stronger than brain signals.
Meta plans to confront these challenges in upcoming research by enhancing the accuracy and reliability of the decoding processes, exploring alternative non-invasive brain imaging techniques that are feasible for everyday use, and developing advanced AI models to better interpret intricate brain signals. The company’s ambition extends to expanding research to encompass a wider range of cognitive processes and exploring potential applications across diverse fields such as healthcare, education, and human-computer interaction.
While further research is necessary before these developments can substantially aid individuals with brain injuries, they pave the way towards AI systems that can learn and reason more similarly to humans. The strides made by Meta and collaborating researchers mark significant progress. As they continue to unravel the complexities of human thought processes and language formation, we move ever closer to a future where technology bridges gaps in human communication and interaction.
