The more AI evolves, the more it resembles the human brain! Meta found the "prefrontal cortex" of the machine. AI scholars and neuroscientists were surprised

Fish and sheep Xiao Xiao From the Aofei temple
qubits | official account QbitAI

You may not believe it , There is one. AI Just proven , The way we deal with voice is the same as The brain The mystery is similar .

Even in structure Can correspond to each other ——

Scientists in AI Directly located on the body “ Visual cortex ”.

This one comes from Meta AI And so on po Out , It immediately exploded on social media . A wave of neuroscientists and AI The researchers went to the onlookers .

LeCun Praise this is “ Excellent work ”： Self supervision Transformer Between stratification and human auditory cortex , It is indeed closely related .

Some netizens took the opportunity to make fun of ：Sorry Marcus , but AGI It's really coming .

however , The research has also aroused the curiosity of some scholars .

For example, Dr. of neuroscience, McGill University Patrick Mineault Ask questions ：

We published in NeurIPS In a paper of , Also tried to fMRI Data and models are linked , But I didn't think there was any relationship between the two .

therefore , What kind of research is this , How does it come to “ This one AI Work like a brain ” The conclusion is ？

AI Learn to work like a human brain

Simply speaking , In this study , Researchers focus on speech processing problems , The self-monitoring model Wav2Vec 2.0 Same as 412 name The brain activity of the volunteers was compared .

this 412 Of the volunteers , Yes 351 People speak English ,28 People speak French ,33 People speak Chinese . The researchers listened to them for about 1 Hours of audio books , And in this process fMRI Their brain activity was recorded .

Model side , Researchers used more than 600 Hours of unlabeled voice training Wav2Vec 2.0.

Corresponding to the volunteer's mother tongue , The model is also divided into English 、 French 、 Three in Chinese , Another one is trained with non speech acoustic scene data set .

Then the models listened to the same audio book of volunteers . The researchers extracted the activation of the model .

Evaluation criteria of relevance , Follow this formula ：

among ,X Activate... For the model ,Y For human brain activity ,W For the standard coding model .

From the results , Self supervised learning does make Wav2Vec 2.0 It produces speech representations similar to those of the brain .

As you can see from the above figure , In the primary and secondary auditory cortex ,AI It clearly predicts brain activity in almost all cortical areas .

The researchers also found that AI Of “ Auditory cortex ”、“ Prefrontal cortex ” Which floor is it on .

The picture shows , Auditory cortex and Transformer The first floor of （ Blue ） Best match , The prefrontal cortex is associated with Transformer The deepest layer of （ Red ） Best match .

Besides , The researchers quantitatively analyzed the differences in human ability to perceive native and non-native phonemes , And with Wav2Vec 2.0 Compare the models .

They found that ,AI Just like human beings , Yes “ mother tongue ” Have a stronger ability to distinguish , such as , The French model is easier to perceive the stimulation from French than the English model .

The above results prove ,600 Hours Self supervised learning , Enough to make Wav2Vec 2.0 Learning specific representations of language —— This is similar to what babies come into contact with in the process of learning to speak “ Data volume ” Quite a .

Need to know , Before DeepSpeech2 The paper holds that , Need at least 10000 Hours Voice data （ It has to be the marked one ）, To build a good set of voice to text （STT） System .

Once again, neuroscience and AI Boundary discussion

For this study , Some scholars think , It did make some new breakthroughs .

for example , From Google brain Jesse Engel call , This research takes the visualization filter to a new level .

Now? , Not only can you see them in “ Pixel space ” What does it look like , Connect them to “ Brain like space ” You can also simulate the shape in ：

For example , front MILA And Google researchers Joseph Viviano Think , This study also proves that fMRI The resting state in （resting-state） Imaging data is meaningful .

But in a discussion , There were also some voices of doubt .

for example , Doctor of neuroscience Patrick Mineault In addition to pointing out that they have done similar research but have not reached a conclusion , Also gave some of their own doubts .

He thinks that , This study does not really prove that it measures “ Voice Processing ” The process of .

Compared to the speed at which people talk ,fMRI The speed of measuring the signal is actually very slow , Therefore, we hastily come to the conclusion that “Wav2vec 2.0 Learning the behavior of the brain ” The conclusion is unscientific .

Of course ,Patrick Mineault He said that he did not deny the view of the study , He himself is “ One of the author's fans ”, But this study should give some more convincing data .

In addition, some netizens think ,Wav2vec And the input of human brain is not the same , One is the processed waveform , But the other is the original waveform .

Regarding this , One of the authors 、Meta AI researcher Jean-Rémi King summary ：

Simulate human level intelligence , There is really a long way to go . But at least for now , We may be on the right path .

Do you think? ？

Address of thesis ：
https://arxiv.org/abs/2206.01685

Reference link ：
[1]https://twitter.com/patrickmineault/status/1533888345683767297
[2]https://twitter.com/JeanRemiKing/status/1533720262344073218
[3]https://www.reddit.com/r/singularity/comments/v6bqx8/toward_a_realistic_model_of_speech_processing_in/
[4]https://twitter.com/ylecun/status/1533792866232934400