Connectionists: short Op-ed to address AI problems

Weng, Juyang weng at msu.edu
Thu Jun 6 23:09:03 EDT 2024


Dear Asim,
   You wrote, "Let’s do one issue at a time. Let’s try symbols first."  This approach misleads you to the wrong track.
   Case 1: neuron level symbols (your position).
   Case 2: area level symbols.
   Case 3: task level symbols.
   They are all dead ends because Asim is the government of the "brain" model.
    For all those Asim knows, it is too expensive to create all symbols for the "brain" model.
    For all those Asim does not know, the model does not know either.
    Deadends!  If you continue this "one issue at a time route," you waste too much time in your life.  This is because the first issue is wrong to consider.
    Best regards,
-John Weng
________________________________
From: Asim Roy <ASIM.ROY at asu.edu>
Sent: Thursday, June 6, 2024 10:06 PM
To: Weng, Juyang <weng at msu.edu>; Stephen José Hanson <jose at rubic.rutgers.edu>; Gary Marcus <gary.marcus at nyu.edu>
Cc: connectionists at mailman.srv.cs.cmu.edu <connectionists at mailman.srv.cs.cmu.edu>
Subject: RE: Connectionists: short Op-ed to address AI problems


Dear John,



Let’s do one issue at a time. Let’s try symbols first. There is plenty of evidence in neurophysiology that one can associate “meaning” to the activation of certain individual cells. As far as I know, all of the brain-related Nobel prizes were about finding “meaning” in the activations of certain single neurons. Here I quote from Wikipedia (Single-unit recording - Wikipedia<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/Single-unit_recording__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNkNR1ZZ9A$>):



  *   1928: One of the earliest accounts of being able to record from the nervous system was by Edgar Adrian<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/Edgar_Adrian__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNnQtx1LXQ$> in his 1928 publication "The Basis of Sensation". In this, he describes his recordings of electrical discharges in single nerve fibers using a Lippmann electrometer<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/Lippmann_electrometer__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNn6lQGgzA$>. He won the Nobel Prize in 1932 for his work revealing the function of neurons.[11]<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/Single-unit_recording*cite_note-11__;Iw!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNk6I8LNhA$>
  *   1957: John Eccles<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/John_Carew_Eccles__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNkFV1ULMA$> used intracellular single-unit recording to study synaptic mechanisms in motoneurons (for which he won the Nobel Prize in 1963).
  *   1959: Studies by David H. Hubel<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/David_H._Hubel__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNkJaa_aew$> and Torsten Wiesel<https://urldefense.com/v3/__https://en.wikipedia.org/wiki/Torsten_Wiesel__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNndrzwVDg$>. They used single neuron recordings to map the visual cortex in unanesthesized, unrestrained cats using tungsten electrodes. This work won them the Nobel Prize in 1981 for information processing in the visual system.



  *   And the work of Mosers and O’Keefe on grid and place cells that won them the Nobel: The 2014 Nobel Prize in Physiology or Medicine - Press release<https://urldefense.com/v3/__https://www.nobelprize.org/prizes/medicine/2014/press-release/__;!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNm67pI7iQ$>. Here’s a quote about the work on place cells:



“Most neuroscientists once doubted that brain activity could be linked with behaviour, but in the late 1960s, O’Keefe began to record signals from individual neurons in the brains of rats moving freely in a box. He put electrodes in the hippocampus and was surprised to find that individual cells fired when the rats moved to particular spots.” Nobel prize for decoding brain’s sense of place | Nature<https://urldefense.com/v3/__https://www.nature.com/articles/514153a*:*:text=Most*20neuroscientists*20once*20doubted*20that*20brain*20activity*20could,fired*20when*20the*20rats*20moved*20to*20particular*20spots.__;I34lJSUlJSUlJSUlJSUlJQ!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNnqimHsFw$>



And then the findings about concept cells (Jennifer Aniston cells), which are single cell recordings. Here’s from Reddy and Thorpe (2014)<https://urldefense.com/v3/__https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.00059/full*B6__;Iw!!HXCxUKc!2pFG0g1tPh-88cfwjJImIxJxtBhaOQ1wWf15ZEUkChi5vUb8q_qEXUDZt7bsQ9QjqzSglNlbw-6x9Q$>: “concept cells have “meaning of a given stimulus in a manner that is invariant to different representations of that stimulus.”



We all try to generalize from data, right. If you examine these findings, the most important feature is that they all found “meaning” in single cell activations. So the most fundamental question for you is: Do you accept these findings and the general conclusion that single cell activations can have meaning? Again, beware that, beyond winning Nobel prizes, much work in neuroscience and other fields follows from these findings.



All the best,

Asim Roy

Professor, Information Systems

Arizona State University

Lifeboat Foundation Bios: Professor Asim Roy<https://urldefense.proofpoint.com/v2/url?u=https-3A__lifeboat.com_ex_bios.asim.roy&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=oDRJmXX22O8NcfqyLjyu4Ajmt8pcHWquTxYjeWahfuw&e=>

Asim Roy | iSearch (asu.edu)<https://urldefense.proofpoint.com/v2/url?u=https-3A__isearch.asu.edu_profile_9973&d=DwMFaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=waSKY67JF57IZXg30ysFB_R7OG9zoQwFwxyps6FbTa1Zh5mttxRot_t4N7mn68Pj&s=jCesWT7oGgX76_y7PFh4cCIQ-Ife-esGblJyrBiDlro&e=>







From: Weng, Juyang <weng at msu.edu>
Sent: Thursday, June 6, 2024 1:09 AM
To: Asim Roy <ASIM.ROY at asu.edu>; Stephen José Hanson <jose at rubic.rutgers.edu>; Gary Marcus <gary.marcus at nyu.edu>; Weng, Juyang <weng at msu.edu>
Cc: connectionists at mailman.srv.cs.cmu.edu
Subject: Re: Connectionists: short Op-ed to address AI problems



Dear Asim,

   You wrote, "We are doing neurosymbolic with image processing – the symbolic stuff on top of a DL model. It also brings in the explanation side."

   Not only DL is misconduct, but symbols are another devil.

   In my IJCNN 2022 paper,

   http://www.cse.msu.edu/~weng/research/20M-IJCNN2022rvsd-cite.pdf<https://urldefense.com/v3/__http:/www.cse.msu.edu/*weng/research/20M-IJCNN2022rvsd-cite.pdf__;fg!!IKRxdwAv5BmarQ!YZcFaLmNraAEJLpxRQGKzKZTVt_nn3J9i52_xG7zhEgKn6ZASf_q59sOFVdSPylt7_NueMymM_EI7GNl$>

   I proved "symbol-free" as one of the 20 million-dollar problems for us to understand human brains.

   The definition of symbols requires a government,  but government-free is one of the 20 million-dollar problems for us to understand human brains.

   Let us consider three cases:

  Case 1:  If a human designs symbols within a network (e.g., LSTM) and assigns the symbols to some individual neurons (e.g., task-specific gates) of the network, this human is a government within the network since he is task-aware.

  Case 2: If a human designs symbols within a network and assigns roles to blocks in a functional block diagram, e.g., [Starzyk10], this human is a government within the network.

  Case 3: In the symbolic AI school, a human programmer designs symbolic representations for a task that is assigned to a computer program or network.  This human is a government within the symbolic AI system since he is task-aware.

   All the 3 cases do not solve the government-free problem.

  I have attached an image that further explains the symbol problem in the same paper.

  Let me know if you still do not agree that the brain must be free from symbols after you read the entire paper.

   By the way, I am surprised that as a mathematician, you still do not understand the Post-Selection misconduct in DL that I raised to you earlier.  Please use your own words to explain Post-Selection and why you can handle explanation using Post-Selection misconduct.

   Best regards,

-John Weng



________________________________

From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> on behalf of Asim Roy <ASIM.ROY at asu.edu<mailto:ASIM.ROY at asu.edu>>
Sent: Wednesday, June 5, 2024 6:49 PM
To: Stephen José Hanson <jose at rubic.rutgers.edu<mailto:jose at rubic.rutgers.edu>>; Gary Marcus <gary.marcus at nyu.edu<mailto:gary.marcus at nyu.edu>>
Cc: connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu> <connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu>>
Subject: Re: Connectionists: short Op-ed to address AI problems



Dear Stephen,



We are doing neurosymbolic with image processing – the symbolic stuff on top of a DL model. It also brings in the explanation side. The results are astounding. We get better performance than a pure DL model. And exploring applications with defense agencies. They are impressed with the results we have so far. So, neurosymbolic is definitely the way forward.



Best,

Asim Roy

Professor, Information Systems

Arizona State University

Asim Roy | ASU Search<https://urldefense.com/v3/__https:/search.asu.edu/profile/9973__;!!HXCxUKc!1ZzSj6Uim5wWu5W-JiBNqp_Cig3tUkK5DgMhDEBYnERP1f-pOAReghJiHzEk3hEHKL31roB_8qivsA$>

Lifeboat Foundation Bios: Professor Asim Roy<https://urldefense.com/v3/__https:/lifeboat.com/ex/bios.asim.roy__;!!HXCxUKc!1ZzSj6Uim5wWu5W-JiBNqp_Cig3tUkK5DgMhDEBYnERP1f-pOAReghJiHzEk3hEHKL31roAnhJU86A$>



From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> On Behalf Of Stephen José Hanson
Sent: Wednesday, June 5, 2024 6:06 AM
To: Gary Marcus <gary.marcus at nyu.edu<mailto:gary.marcus at nyu.edu>>
Cc: connectionists at mailman.srv.cs.cmu.edu<mailto:connectionists at mailman.srv.cs.cmu.edu>
Subject: Re: Connectionists: short Op-ed to address AI problems



Dear Flabbergasted:

Thankyou, I endeavor to provide short but useful commentary that could be considered a "work of art".  Graci!

Now either my memory is failing since 2017(not impossible), or you are smoothing over a time series of claims that are actually like a seesaw.

I think if we just rewind some of the connectionist comments; it would be clear, in fact, for example, you had a long series of comments with Geoff that seemed to indicate you were being misreprented as well.  Your complaints have always be around the fact that DL-AI has false alarms (and to be fair other problems)   And sometimes pretty odd-ones.  LLMs human and non-human errors are even more interesting.  The fact that they seem to grow circuits in the attention-heads is gobsmacking!   I thought then and think now you are complaining about peas under a very thick mattress (oh-oh,  metaphors now- I may have opened pandora's box.)

But I will go look at the budding NeuroSymbolic paper you mentioned, but I have my doubts that the statistical bias is equivalent with the architectually simplistic LLMs.  Nonetheless, I have not read it.

I will also make a  coarse  timeline of your comments since 2017, but anyone out there that would like to help, greatly appreciated!

Best,

Stephen

On 6/5/24 8:41 AM, Gary Marcus wrote:

Wow, Stephen, you have outdone yourself. This note is a startling mixture of rude, condescending, inaccurate, and uninformed. A work of art!



To correct four misunderstandings:

1. Yes, my essay was written before LLMs were popular (though around the time Transformers were proposed as it happens). It was however precisely “  a moonshot idea, that doesn't involve leaving the blackbox in the hands of corporate types who value profits over knowledge.” Please read what I wrote. It’s one page, linked below, and you obviously couldn’t be bothered,. (Parenthetically, I was one of the first people to warn that OpenAI was likely to be problematic,  and have done so repeatedly at my Substack.)

2. My argument throughout (back to 2012, in the New Yorker, 2018 in my Deep Learning: A Critical Appraisal, etc) has been that deep learning has some role but cannot solve all things, and that it would be not reliable on its own. In 2019 onwards I emphasized many of the social problems that arise from relying on such unreliable architectures. I have never wavered from any of that. (Again, please read my work before so grossly distorting it.) Unreliable systems that are blind to truth and values can cause harm (bias), be exploited (to create disinformation), etc. There is absolutely no contradiction there, as I have explained numerous times in my writings.

3. It’s truly rude to dismiss an entire field as “flotsam and jetsam”,  and you obviously aren’t following the neurosymbolic literature, e.g., you must have missed DeepMind’s neurosymbolic AlphaGeometry paper, in Nature, with its state of the art results, beating pure neural nets.

4. Again, nothing has changed about my view; your last remark is gratuitous and based on a misunderstanding.



Truly flabbergasted,

Gary



On Jun 5, 2024, at 05:18, Stephen José Hanson <jose at rubic.rutgers.edu><mailto:jose at rubic.rutgers.edu> wrote:



Gary, this was before the LLM discovery.   Pierre is proposing a moonshot idea, that doesn't involve leaving the blackbox in the hands of corporate types who value profits over knowledge.  OPENAI seems to be flailing and having serious safety and security issues.  It certainly could be recipe for diaster.

Frankly your views have been all over the place.  DL doesn't work, DL could work but should be merged with the useless flotsam and jetsam from GOFAI over the last 50 years, and now they are too dangerous because they work but they are unreliable, like most humans.

Its hard to know what views of yours to take seriously as they seem change so rapidly.

Cheers

Stephen

On 6/4/24 9:53 AM, Gary Marcus wrote:

I would just point out that I first made this suggestion [CERN for AI] in the New York Times in 2017, and several others have since. There is some effort ongoing to try to make it happen, if you search you will see.



<30gray-facebookJumbo.jpg>

Opinion | Artificial Intelligence Is Stuck. Here’s How to Move It Forward. (Gift Article)<https://urldefense.proofpoint.com/v2/url?u=https-3A__www.nytimes.com_2017_07_29_opinion_sunday_artificial-2Dintelligence-2Dis-2Dstuck-2Dheres-2Dhow-2Dto-2Dmove-2Dit-2Dforward.html-3Funlocked-5Farticle-5Fcode-3D1.xE0.mcIz.lT-5FK7BZdonGJ-26smid-3Dnytcore-2Dios-2Dshare-26referringSource-3DarticleShare-26u2g-3Di-26sgrp-3Dc-2Dcb&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=fwBsbQ5xjEJFDg3c0iXuOBcr84mxEGxR0cEG4-hstVM8dJNyq3HVvpCACElUGWT2&s=TGZDkK1TsB_rNyjmal5jG1694upjB2JDhtj3UOe4Cws&e=>

nytimes.com<https://urldefense.proofpoint.com/v2/url?u=https-3A__www.nytimes.com_2017_07_29_opinion_sunday_artificial-2Dintelligence-2Dis-2Dstuck-2Dheres-2Dhow-2Dto-2Dmove-2Dit-2Dforward.html-3Funlocked-5Farticle-5Fcode-3D1.xE0.mcIz.lT-5FK7BZdonGJ-26smid-3Dnytcore-2Dios-2Dshare-26referringSource-3DarticleShare-26u2g-3Di-26sgrp-3Dc-2Dcb&d=DwMGaQ&c=slrrB7dE8n7gBJbeO0g-IQ&r=wQR1NePCSj6dOGDD0r6B5Kn1fcNaTMg7tARe7TdEDqQ&m=fwBsbQ5xjEJFDg3c0iXuOBcr84mxEGxR0cEG4-hstVM8dJNyq3HVvpCACElUGWT2&s=TGZDkK1TsB_rNyjmal5jG1694upjB2JDhtj3UOe4Cws&e=>





On Jun 3, 2024, at 22:58, Baldi,Pierre <pfbaldi at ics.uci.edu><mailto:pfbaldi at ics.uci.edu> wrote:


I would appreciate feedback from this group,especially dissenting feedback,  on the attached Op-ed. You can send it to my personal email which you can find on my university web site if you prefer. The basic idea is simple:

IF for scientific, security, or other societal reasons we want academics to develop and study the most advanced forms of AI, I can see only one solution:  create  a national or international effort around the largest data/computing center on Earth with a CERN-like structure comprising permanent staff, and 1000s of affiliated academic laboratories. There are many obstacles, but none is completely insurmountable if we wanted to.

Thank you.

Pierre




<AI-CERN-Baldi2024FF.pdf>

--

Stephen José Hanson

Professor of Psychology

Director of RUBIC

Member of Exc Comm RUCCS

--

Stephen José Hanson

Professor of Psychology

Director of RUBIC

Member of Exc Comm RUCCS
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20240607/641507fd/attachment.html>


More information about the Connectionists mailing list