Scientists at the University of Texas at Austin have built a “semantic brain decoder” that can predict what someone is thinking based on their brain activity. The decoder learns to match language patterns with different types of brain activity using MRI brain scans of individuals listening to 16 hours of podcasts. It then makes “guesses” of what the listener is thinking and ranks them according to how closely they resemble the person’s brain activity. While it does not have the capability to read thoughts against one’s will, it is intended to aid individuals who cannot speak due to stroke or other medical emergencies.
The thought-based predictions made by the decoder translated the general meaning of the thoughts rather than the literal translation, indicating that it captured the essence of viewers’ thinking. The decoder’s predictions were compared to audio recordings of the audio or a text version of the story the viewer conjured up. During the tests, participants tried to confound the decoder by remaining obstinate, which resulted in the decoder producing unintelligible sounds.
Huth and his team hypothesized about the possibility of the decoder technology being utilized for sinister purposes such as lie detection or against the victim’s will to access and reveal their stored memories. As a result, they conducted tests to investigate what the decoder could not do. One examination involved participants attempting to disrupt the decoder by individually naming an increasing number of animals while listening to a story. The decoder was unable to handle this complexity and produced gibberish. Another experiment sought to understand the decoder’s ability to read the thoughts of an individual who had not yet listened to 16 hours of podcasts, as this was the data used to train the decoder. This study was likewise ineffective.
One issue that arises with AI is that the data used to train the model can have human biases. For example, if it can relate doctors to men and nurses to women, it can cause errors in prediction. In the case of the brain decoder, participants listen to podcasters of various genders, ethnicities, and sexualities. The researchers, however, have had less oversight over the data utilized for training the GPT models, which may reduce the decoder’s prediction accuracy but may not indicate partiality.
The brain decoder is intended to assist those who cannot speak following a stroke or other medical occurrence, but it currently needs an expensive and immobile MRI machine to work. The decoding process requires hours of data processing. As a result, Huth’s team is examining alternative and less expensive methods for determining brain activity, such as functional near-infrared spectroscopy (fNIRS). While the decoder seems to have tapped into how humans interpret language to understand the world, it cannot analyze memories. The viewer linking “hamburger,” “fries,” and “milkshake” with a specific brain activity pattern is only one aspect of how we experience the world, because human cognitive capacity is more complex than simply predicting the next word.
Editor’s remarks: This article explains the innovative “semantic brain decoder” technology created at the University of Texas, which can guess people’s thoughts based on their brain activity but is not yet capable of reading memories. This technology is not currently used maliciously without subjects’ consent, according to researchers. Nevertheless, it is critical to investigate potential privacy issues, given that the future utilization of this technology is unknown. As a result, it is critical to conduct additional research on potential biases and to explore strategies for ensuring that privacy is not breached. Visit the GPT News Room for more fascinating tech articles like this one.
Source link
from GPT News Room https://ift.tt/p70mMiT
No comments:
Post a Comment