The artificial intelligence model illustrates that how the human brain understands the language

Nov, 2021 - By SMI

The model is used to predict the next word and complete tasks that appear to need some level of conceptual understanding, such as answering questions, text summarizing, and storytelling.

Artificial intelligence models of language have been exceedingly good at some tasks in recent years. They excel, most significantly, at guessing the next word in a text string, this technology assists search engines and messaging apps in estimating the next phrase you will enter. The newer versions of predictive language models appear to be learning something about language's underlying message.

Such models were created to optimize performance for the specific purpose of text prediction, rather than to emulate how the human brain does this activity or perceives language. However, according to new research by MIT neuroscientists, the basic function of these models is similar to that of language-processing areas in the human brain. Computer models that do well on other sorts of language tasks lack this closeness to the human brain, indicating that the human brain may employ next-word forecasting to promote language processing.

The MIT researchers adopted a similar strategy in the latest study that compared language-processing centers in the brain with language-processing algorithms. The study examined 43 different language models, including several others that are designed to predict the following word. These include the GPT-3 (Generative Pre-trained Transformer 3) model, which can generate language comparable to what a person might produce when given a stimulus. Other models were created to carry out the specific task given, such as filling in a blank in a phrase.

The researchers tested the behavior of the network's nodes as each model was provided with a set of words. They then matched these patterns to brain activity, which was monitored in people doing three linguistic tasks: listening to the story, reading phrases one at a time, and reading statements with one word disclosed at a time. These human datasets comprised functional magnetic resonance imaging (fMRI) data and intracranial electrocorticographic readings taken in epilepsy patients receiving brain surgery. They discovered that the best-performing next-word prediction models had patterns of activity that were very similar to those observed in the human brain. Activity in those identical models was also substantially associated with human behavioral measurements such as how quickly people could read the text.

The researchers intend to construct versions of these language processing algorithms to explore how tiny changes in their architecture influence their work and capacity to suit human brain data.