The neural architecture of language

[Michael DeBellis]

I came across the following article that seemed to make some pretty strong claims for machine learning approaches as a model for human language comprehension: https://interestingengineering.com/ai-mimicking-the-brain-on-its-own 

I've found that it is important to look at the actual research because things often get distorted and exaggerated when presented in the popular press. The actual journal paper is here: https://www.pnas.org/content/118/45/e2105646118  However, that seems to require an academic login which I don't have but I found an earlier version of the paper that was freely available and more or less the same: https://www.biorxiv.org/content/10.1101/2020.06.26.174482v3.full.pdf+html  I've been trying to read this but I just can't get it. What really has me confused is that they seem to be saying that the Machine Learning models predict both NLP facts (e.g., given the first words of a sentence predicting the next word) but that those same models also predict (or correlate to or ???) neural states in people's brains when they look at similar language input. But I feel that I must be misunderstanding that because even if the neural net was analogous to a language processing neural net in the brain there is no way that fMRI or other measures of brain activity come anywhere close to measuring things like which neurons connect and are firing and inhibiting/triggering which other neurons (not to mention many of the ML models aren't even neural nets). I'm just wondering if anyone can take a look at this and provide some high level summary of what the claim is?

Michael

[James Davenport]

I have the PNAS article, which I’ll try to look at.  Note that there has been, at least, title drift between bioarxiv and published:

Bioarxiv: Artificial Neural Networks Accurately Predict Language Processing in the Brain.

PNAS: The neural architecture of language: Integrative modeling converges on predictive processing.

 

From: ontolo...@googlegroups.com <ontolo...@googlegroups.com> On Behalf Of Michael DeBellis
Sent: 24 December 2021 00:59
To: ontolog-forum <ontolo...@googlegroups.com>
Subject: [ontolog-forum] The neural architecture of language

 

CAUTION:  This email came from outside of the University. To keep your account safe, only click on links and open attachments if you know the person who sent the email, or you expected to receive this communication.

 

I came across the following article that seemed to make some pretty strong claims for machine learning approaches as a model for human language comprehension: https://interestingengineering.com/ai-mimicking-the-brain-on-its-own 

 

I've found that it is important to look at the actual research because things often get distorted and exaggerated when presented in the popular press. The actual journal paper is here: https://www.pnas.org/content/118/45/e2105646118  However, that seems to require an academic login which I don't have but I found an earlier version of the paper that was freely available and more or less the same: https://www.biorxiv.org/content/10.1101/2020.06.26.174482v3.full.pdf+html  I've been trying to read this but I just can't get it. What really has me confused is that they seem to be saying that the Machine Learning models predict both NLP facts (e.g., given the first words of a sentence predicting the next word) but that those same models also predict (or correlate to or ???) neural states in people's brains when they look at similar language input. But I feel that I must be misunderstanding that because even if the neural net was analogous to a language processing neural net in the brain there is no way that fMRI or other measures of brain activity come anywhere close to measuring things like which neurons connect and are firing and inhibiting/triggering which other neurons (not to mention many of the ML models aren't even neural nets). I'm just wondering if anyone can take a look at this and provide some high level summary of what the claim is?

 

Michael

--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info/
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/461bb270-3dcb-45b3-859b-35588c4ee414n%40googlegroups.com.

[Michael DeBellis]

James, thanks for pointing that out about the title drift and for looking at it. Any chance you could share the PNAS article? I think you can attach documents in comments (or if I just missed the link for the PDF if you could include that). Just a bit more about why I want to understand the article: I'm very skeptical of claims that machine learning models provide much insight into how humans process language. I think that there is a difference between AI in the engineering sense and AI in the sense of using computer models to understand the mind/brain and just because ML systems do a better job (in general) than semantic/symbolic systems in terms of engineering solutions to NLP doesn't mean that the way they work is a model for how the human mind works. 

However, just for that reason I want to understand their claims better (a good scientists has to realize he may be wrong and must try to understand different points of view). I'm also trying to educate myself more on Machine Learning in general. I need to go and finish a Coursera class by Andrew Ng. But I also find that sometimes just immersing myself into the "deep end" so to speak can result in a lot of learning. Just from looking up things like "embedding" and understanding the context of the article I've come to understand at least that idea better. I appreciate any insight you can share. Merry Christmas.

Michael 

[James Davenport]

I may be wrong here, but I THINK what they are claiming is that, using the same structures as standard NLP ANNs, in particular GPT-2, they can build ANNs which do well at predicting the signals they CAN measure in the brain.

The key sentence for me (page 2) is this.

Specifically, we examined the relationships between 43 diverse

state-of-the-art ANN language models (henceforth “models”)

across three neural language comprehension datasets (two fMRI,

one electrocorticography [ECoG]), as well as behavioral signatures

of human language processing in the form of self-paced reading

times, and a range of linguistic functions assessed via standard

engineering tasks from NLP.

To view this discussion on the web visit https://groups.google.com/d/msgid/ontolog-forum/LO2P265MB44006E5CE6D01B8AAC3580BFE07F9%40LO2P265MB4400.GBRP265.PROD.OUTLOOK.COM.

[Michael DeBellis]

> using the same structures as standard NLP ANNs, in particular GPT-2, they can build ANNs which 

>do well at predicting the signals they CAN measure in the brain  , in particular GPT-2, 

>they can build ANNs which do well at predicting the signals they CAN measure in the brain

Thanks, that is what I thought as well. The question I have is what are the specifics of " using the same structures as standard NLP ANNs...  they can build ANNs which do well at predicting the signals they CAN measure in the brain" I.e., when they "use the same structures" what exactly does that mean and how specific is that to predicting brain states. I.e., could you "use the same structure" to recognize other kinds of signals in noisy data? In fact they talk about similar work done in visual recognition. So would it be possible to "use the same structures" to "build ANNs which wo well at predicting various shapes in a stream of visual data? When they build the ANNs to predict brain states, how much of the original model remains vs. how much was it altered? I also think there is a bit of a difference between visual and language processing in animals because primates and other mammals process visual information very much the way humans do so we know much more about visual processing because we can do experiments on other animals we can't do on humans. But most people agree that the way humans process language is unique. So when he talks about certain processes that are localized in brains for visual processing we have a much better idea of exactly what is being done where (e.g., the Lateral Geniculate Nucleus is similar to a router for sense data, the various levels of the visual cortex process things from low level like edges to high levels like faces going from levels 1 to higher levels). So I forget exactly what part of visual processing they talk about but it seems more realistic to say that an ANN can replicate a certain type of visual processing done in a certain part of the brain where as for language what we know (Wernicke's and Broca's areas) is not nearly as well understood and there is still a lot of disagreement about how local various linguistic functions are and where they take place. 

I'm just a neophyte when it comes to brain science so if anything I said was wrong or you have further thoughts please let me know. I still don't understand the paper completely but trying to figure it out helped me understand some ML concepts that I've heard about in the past but never really understood. Thanks for your help.

Michael