Can't help but want to answer?

Xiao Xiao from

In fact, your brain is like an auto-completion machine, guessing what other people are going to say all the time

Not only that, but the brain is constantly comparing the information we get from the outside world (images, sounds, text, etc.

Based on this theory, researchers have come up with many NLP models, including speech recognition, automatic completion, etc.

So, is it because the human brain and AI's "brain supplement" method are different?

Now, a group of scientists from the Max Planck Institute and Radboud University have decided to directly measure the brain waves that people emit when they are "brain tonic"

After analyzing these brain wave signatures, they discovered more detailed reasons behind it

Collect brain 'reactions' when listening to books

Scientists have previously discovered that the brain makes predictions about external events to guide itself in processing information

However, there is no clear conclusion on what conditions (parts of speech/phonemes/semantics, etc.

For example, how words and semantics affect brain predictions, and whether predicted words affect semantics, is unclear

Therefore, the scientists came up with a method: let the volunteers listen to the audiobook, observe their brain responses, and use GPT-2 to conduct a quantitative analysis of the predicted content and compare the predicted results

They first found 19 volunteers, measured and recorded their brain waves (EEG) for 1 hour; then, collected 9 hours of magnetoencephalography (MEG) of 3 volunteers to locate the neural activity in specific locations in the brain

Among them, the EEG group listened to "The Old Man and the Sea" for one hour, each time was 180 seconds and played 20 times in total; the magnetoencephalography group listened to "Sherlock Holmes" for nine hours, one hour each time Asked to answer specific questions between intervals

Subsequently, these data were analyzed from different dimensions, and several conclusions were drawn

Prioritize "more advanced" forecast conditions

The researchers first verified that the brain's predictions are produced continuously

In the process, scientists have also discovered similarities between AI predictions and the human brain

For example, if the predictions don't match reality (guess wrong), the brain gets "surprised" and becomes more active than ever

They then divided the content of the audiobook into several levels, including "levels" of parts of speech (verb/noun/adjective), phonemes (phonetic symbols), and semantics (predicted with GPT-2)
.

Then, MRI restored the "visualization map" of the brain as it made different levels of prediction
.

The results show that different levels of the brain interact with each other when making predictions using parts of speech, semantics and phonemes
.

That is, more "advanced" prediction conditions affect lower-priority prediction conditions, such as semantic (context) effects on phoneme prediction
.

For example, in the sentence "They say his father was a _____.
", predicting the next syllable only by phoneme is different from predicting by context
.

If you only use phonemes and frequencies to make predictions, there is a high probability that the word phonetic transcription will start with /fɔ/, but combined with the context, it is easier to make "fisherman", that is, word predictions starting with /fi/
.

In this way, it may also be possible to increase the accuracy of AI such as speech recognition or autocompletion by combining several levels of prediction conditions
.

about the author

The first is Micha Heilbron, who is currently at the University of Nijmegen and the Max Planck Institute.
His research direction is cognitive and computational neuroscience, and he is interested in the role of generative models in language processing
.

Kristijan Armeni, currently a postdoctoral fellow at Johns Hopkins University, graduated with a Ph.
D.
from the University of Najmegen, with research interests in Natural Language Processing and Cognitive Neuroscience
.

Several authors who also participated in the study, Jan-Mathijs Schoffelen, Peter Hagoort, and Floris P.
de Lange, are also from the University of Nejmegen and the Max Planck Institute, respectively
.

For this study, some netizens put forward their own point of view, that the brain itself is a mechanism that makes decisions and actions based on a feedback loop
.

But there are also netizens who, from their own feelings, seem to often make mistakes as the "prediction machine" of the brain
.

Some netizens put forward their views on this:

After all, the brain is a survival machine, not a "truth detector.

"
If lying promotes health, your brain will choose to lie
.

Paper address:
link:
[1]https://news.
ycombinator.
com/item?id=32395840
[2] Finish-

"Artificial Intelligence", "Smart Car" WeChat community invites you to join!

Welcome friends who are interested in artificial intelligence and smart cars to join us, communicate and learn from AI practitioners, and do not miss the latest industry development & technological progress
.

ps.
Please be sure to note your name-company-position when adding friends~

Click here👇 to follow me, remember to mark the star~

"Share", "Like" and "Watching" in one click

See you every day at the forefront of science and technology~