Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc.

Sun Nov 7 14:01:30 EST 2021

Yoshua,

I am indeed feeling that I can have the cake and eat it too. Accepting the fact that neural activations in the brain have “meaning and interpretation” is a huge step forward for the field. I would conjecture that it opens the door to new theories in cognitive and neuro sciences. You are definitely crossing the red line and that’s great. Can we have references to some of your papers?

By the way, I think I understand what you mean by disentangling. There are probably simpler ways to disentangle and get to Explainable AI. But please send us the references.

Best,
Asim

From: Yoshua Bengio <yoshua.bengio at mila.quebec>
Sent: Sunday, November 7, 2021 8:55 AM
To: Asim Roy <ASIM.ROY at asu.edu>
Cc: Adam Krawitz <akrawitz at uvic.ca>; connectionists at cs.cmu.edu; Juyang Weng <juyang.weng at gmail.com>
Subject: Re: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc.

Asim,

You can have your cake and eat it too with modular neural net architectures. You still have distributed representations but you have modular specialization. Many of my papers since 2019 are on this theme. It is consistent with the specialization seen in the brain, but keep in mind that there is a huge number of neurons there, and you still don't see single grand-mother cells firing alone, they fire in a pattern that is meaningful both locally (in the same region/module) and globally (different modules cooperate and compete according to the Global Workspace Theory and Neural Workspace Theory which have inspired our work). Finally, our recent work on learning high-level 'system-2'-like representations and their causal dependencies seeks to learn 'interpretable' entities (with natural language) that will emerge at the highest levels of representation (not clear how distributed or local these will be, but much more local than in a traditional MLP). This is a different form of disentangling than adopted in much of the recent work on unsupervised representation learning but shares the idea that the "right" abstract concept (related to those we can name verbally) will be "separated" (disentangled) from each other (which suggests that neuroscientists will have an easier time spotting them in neural activity).

-- Yoshua

I'm overwhelmed by emails, so I won't be able to respond quickly or directly. Please write to my assistant in case of time sensitive matter or if it entails scheduling: julie.mongeau at mila.quebec<mailto:julie.mongeau at mila.quebec>

Le dim. 7 nov. 2021, à 01 h 46, Asim Roy <ASIM.ROY at asu.edu<mailto:ASIM.ROY at asu.edu>> a écrit :
Over a period of more than 25 years, I have had the opportunity to argue about the brain in both public forums and private discussions. And they included very well-known scholars such as Walter Freeman (UC-Berkeley), Horace Barlow (Cambridge; great grandson of Charles Darwin), Jay McClelland (Stanford), Bernard Baars (Neuroscience Institute), Christof Koch (Allen Institute), Teuvo Kohonen (Finland) and many others, some of whom are on this list. And many became good friends through these debates.

We argued about many issues over the years, but the one that baffled me the most was the one about localist vs. distributed representation. Here’s the issue. As far as I know, although all the Nobel prizes in the field of neurophysiology – from Hubel and Wiesel (simple and complex cells) and Moser and O’Keefe (grid and place cells) to the current one on discovery of temperature and touch sensitive receptors and neurons - are about finding “meaning” in single or a group of dedicated cells, the distributed representation theory has yet to explain these findings of “meaning.” Contrary to the assertion that the field is open-minded, I think most in this field are afraid the to cross the red line.

Horace Barlow was the exception. He was perhaps the only neuroscientist who was willing to cross the red line and declare that “grandmother cells will be found.” After a debate on this issue in 2012, which included Walter Freeman and others, Horace visited me in Phoenix at the age of 91 for further discussion.

If the field is open minded, would love to hear how distributed representation is compatible with finding “meaning” in the activations of single or a dedicated group of cells.

Asim Roy
Professor, Arizona State University
Lifeboat Foundation Bios: Professor Asim Roy<https://urldefense.com/v3/__https:/lifeboat.com/ex/bios.asim.roy__;!!IKRxdwAv5BmarQ!JYoK0hORlllDPMK5nxG1MV8TRdHc4uGvWM3awogw4qslieKdtCnnX7G9gvkI0Xg$>

From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> On Behalf Of Adam Krawitz
Sent: Friday, November 5, 2021 10:01 AM
To: connectionists at cs.cmu.edu<mailto:connectionists at cs.cmu.edu>
Subject: Re: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc.

Tsvi,

I’m just a lurker on this list, with no skin in the game, but perhaps that gives me a more neutral perspective. In the spirit of progress:

  1.  If you have a neural network approach that you feel provides a new and important perspective on cognitive processes, then write up a paper making that argument clearly, and I think you will find that the community is incredibly open to that. Yes, if they see holes in the approach they will be pointed out, but that is all part of the scientific exchange. Examples of this approach include: Elman (1990) Finding Structure in Time, Kohonen (1990) The Self-Organizing Map, Tenenbaum et al. (2011) How to Grow a Mind: Statistics, Structure, and Abstraction (not neural nets, but a “new” approach to modelling cognition). I’m sure others can provide more examples.
  2.  I’m much less familiar with how things work on the applied side, but I have trouble believing that Google or anyone else will be dismissive of a computational approach that actually works. Why would they? They just want to solve problems efficiently. Demonstrate that your approach can solve a problem more effectively (or at least as effectively) as the existing approaches, and they will come running. Examples of this include: Tesauro’s TD-Gammon, which was influential in demonstrating the power of RL, and LeCun et al.’s convolutional NN for the MNIST digits.

Clearly communicate the novel contribution of your approach and I think you will find a receptive audience.

Thanks,
Adam

From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> On Behalf Of Tsvi Achler
Sent: November 4, 2021 9:46 AM
To: gary at ucsd.edu<mailto:gary at ucsd.edu>
Cc: connectionists at cs.cmu.edu<mailto:connectionists at cs.cmu.edu>
Subject: Re: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc.

Lastly Feedforward methods are predominant in a large part because they have financial backing from large companies with advertising and clout like Google and the self-driving craze that never fully materialized.

Feedforward methods are not fully connectionist unless rehearsal for learning is implemented with neurons.  That means storing all patterns, mixing them randomly and then presenting to a network to learn.  As far as I know, no one is doing this in the community, so feedforward methods are only partially connectionist.  By allowing popularity to predominate and choking off funds and presentation of alternatives we are cheating ourselves from pursuing other more rigorous brain-like methods.

Sincerely,
-Tsvi

On Tue, Nov 2, 2021 at 7:08 PM Tsvi Achler <achler at gmail.com<mailto:achler at gmail.com>> wrote:
Gary- Thanks for the accessible online link to the book.

I looked especially at the inhibitory feedback section of the book which describes an Air Conditioner AC type feedback.
It then describes a general field-like inhibition based on all activations in the layer.  It also describes the role of inhibition in sparsity and feedforward inhibition,

The feedback described in Regulatory Feedback is similar to the AC feedback but occurs for each neuron individually, vis-a-vis its inputs.
Thus for context, regulatory feedback is not a field-like inhibition, it is very directed based on the neurons that are activated and their inputs.  This sort of regulation is also the foundation of Homeostatic Plasticity findings (albeit with changes in Homeostatic regulation in experiments occurring in a slower time scale).  The regulatory feedback model describes the effect and role in recognition of those regulated connections in real time during recognition.

I would be happy to discuss further and collaborate on writing about the differences between the approaches for the next book or review.

And I want to point out to folks, that the system is based on politics and that is why certain work is not cited like it should, but even worse these politics are here in the group today and they continue to very strongly influence decisions in the connectionist community and holds us back.

Sincerely,
-Tsvi

On Mon, Nov 1, 2021 at 10:59 AM gary at ucsd.edu<mailto:gary at ucsd.edu> <gary at eng.ucsd.edu<mailto:gary at eng.ucsd.edu>> wrote:
Tsvi - While I think Randy and Yuko's book <https://urldefense.com/v3/__https:/www.amazon.com/dp/0262650541/__;!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7U4snDyk0$> is actually somewhat better than the online version (and buying choices on amazon start at $9.99), there is an online version.<https://urldefense.com/v3/__https:/compcogneuro.org/__;!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7UH2qn4go$>
Randy & Yuko's models take into account feedback and inhibition.

On Mon, Nov 1, 2021 at 10:05 AM Tsvi Achler <achler at gmail.com<mailto:achler at gmail.com>> wrote:
Daniel,

Does your book include a discussion of Regulatory or Inhibitory Feedback published in several low impact journals between 2008 and 2014 (and in videos subsequently)?
These are networks where the primary computation is inhibition back to the inputs that activated them and may be very counterintuitive given today's trends.  You can almost think of them as the opposite of Hopfield networks.

I would love to check inside the book but I dont have an academic budget that allows me access to it and that is a huge part of the problem with how information is shared and funding is allocated. I could not get access to any of the text or citations especially Chapter 4: "Competition, Lateral Inhibition, and Short-Term Memory", to weigh in.

I wish the best circulation for your book, but even if the Regulatory Feedback Model is in the book, that does not change the fundamental problem if the book is not readily available.

The same goes with Steve Grossberg's book, I cannot easily look inside.  With regards to Adaptive Resonance I dont subscribe to lateral inhibition as a predominant mechanism, but I do believe a function such as vigilance is very important during recognition and Adaptive Resonance is one of a very few models that have it.  The Regulatory Feedback model I have developed (and Michael Spratling studies a similar model as well) is built primarily using the vigilance type of connections and allows multiple neurons to be evaluated at the same time and continuously during recognition in order to determine which (single or multiple neurons together) match the inputs the best without lateral inhibition.

Unfortunately within conferences and talks predominated by the Adaptive Resonance crowd I have experienced the familiar dismissiveness and did not have an opportunity to give a proper talk. This goes back to the larger issue of academic politics based on small self-selected committees, the same issues that exist with the feedforward crowd, and pretty much all of academia.

Today's information age algorithms such as Google's can determine relevance of information and ways to display them, but hegemony of the journal systems and the small committee system of academia developed in the middle ages (and their mutual synergies) block the use of more modern methods in research.  Thus we are stuck with this problem, which especially affects those that are trying to introduce something new and counterintuitive, and hence the results described in the two National Bureau of Economic Research articles I cited in my previous message.

Thomas, I am happy to have more discussions and/or start a different thread.

Sincerely,
Tsvi Achler MD/PhD

On Sun, Oct 31, 2021 at 12:49 PM Levine, Daniel S <levine at uta.edu<mailto:levine at uta.edu>> wrote:
Tsvi,

While deep learning and feedforward networks have an outsize popularity, there are plenty of published sources that cover a much wider variety of networks, many of them more biologically based than deep learning.  A treatment of a range of neural network approaches, going from simpler to more complex cognitive functions, is found in my textbook Introduction to Neural and Cognitive Modeling (3rd edition, Routledge, 2019).  Also Steve Grossberg's book Conscious Mind, Resonant Brain (Oxford, 2021) emphasizes a variety of architectures with a strong biological basis.

Best,

Dan Levine
________________________________
From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu<mailto:connectionists-bounces at mailman.srv.cs.cmu.edu>> on behalf of Tsvi Achler <achler at gmail.com<mailto:achler at gmail.com>>
Sent: Saturday, October 30, 2021 3:13 AM
To: Schmidhuber Juergen <juergen at idsia.ch<mailto:juergen at idsia.ch>>
Cc: connectionists at cs.cmu.edu<mailto:connectionists at cs.cmu.edu> <connectionists at cs.cmu.edu<mailto:connectionists at cs.cmu.edu>>
Subject: Re: Connectionists: Scientific Integrity, the 2021 Turing Lecture, etc.

Since the title of the thread is Scientific Integrity, I want to point out some issues about trends in academia and then especially focusing on the connectionist community.

In general analyzing impact factors etc the most important progress gets silenced until the mainstream picks it up Impact Factiors in novel research www.nber.org/.../working_papers/w22180/w22180.pdf<https://urldefense.com/v3/__https:/nam12.safelinks.protection.outlook.com/?url=https*3A*2F*2Fwww.nber.org*2Fsystem*2Ffiles*2Fworking_papers*2Fw22180*2Fw22180.pdf*3Ffbclid*3DIwAR1zHhU4wmkrHASTaE-6zwIs6gI9-FxZcCED3BETxUJlMsbN_2hNbmJAmOA&data=04*7C01*7Clevine*40uta.edu*7Cb1a267e3b6a64ada666208d99ca37f6d*7C5cdc5b43d7be4caa8173729e3b0a62d9*7C1*7C0*7C637713048300122043*7CUnknown*7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0*3D*7C1000&sdata=9o*2FzcYY8gZVZiAwyEL5SVI9TEzBWfKf7nfhdWWg8LHU*3D&reserved=0__;JSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUl!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7UD9hRGNg$>  and often this may take a generation https://www.nber.org/.../does-science-advance-one-funeral...<https://urldefense.com/v3/__https:/nam12.safelinks.protection.outlook.com/?url=https*3A*2F*2Fwww.nber.org*2Fdigest*2Fmar16*2Fdoes-science-advance-one-funeral-time*3Ffbclid*3DIwAR1Lodsf1bzje-yQU9DvoZE2__S6R7UPEgY1_LxZCSLdoAYnj-uco0JuyVk&data=04*7C01*7Clevine*40uta.edu*7Cb1a267e3b6a64ada666208d99ca37f6d*7C5cdc5b43d7be4caa8173729e3b0a62d9*7C1*7C0*7C637713048300132034*7CUnknown*7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0*3D*7C1000&sdata=DgxnJTT7MsN5KCzZlA7VAHKrHXVsRsYhopJv0FCwbtw*3D&reserved=0__;JSUlJSUlJSUlJSUlJSUlJSUlJSUl!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7UapVS1t0$>  .

The connectionist field is stuck on feedforward networks and variants such as with inhibition of competitors (e.g. lateral inhibition), or other variants that are sometimes labeled as recurrent networks for learning time where the feedforward networks can be rewound in time.

This stasis is specifically occuring with the popularity of deep learning.  This is often portrayed as neurally plausible connectionism but requires an implausible amount of rehearsal and is not connectionist if this rehearsal is not implemented with neurons (see video link for further clarification).

Models which have true feedback (e.g. back to their own inputs) cannot learn by backpropagation but there is plenty of evidence these types of connections exist in the brain and are used during recognition. Thus they get ignored: no talks in universities, no featuring in "premier" journals and no funding.

But they are important and may negate the need for rehearsal as needed in feedforward methods.  Thus may be essential for moving connectionism forward.

If the community is truly dedicated to brain motivated algorithms, I recommend giving more time to networks other than feedforward networks.

Video: https://www.youtube.com/watch?v=m2qee6j5eew&list=PL4nMP8F3B7bg3cNWWwLG8BX-wER2PeB-3&index=2<https://urldefense.com/v3/__https:/nam12.safelinks.protection.outlook.com/?url=https*3A*2F*2Fwww.youtube.com*2Fwatch*3Fv*3Dm2qee6j5eew*26list*3DPL4nMP8F3B7bg3cNWWwLG8BX-wER2PeB-3*26index*3D2&data=04*7C01*7Clevine*40uta.edu*7Cb1a267e3b6a64ada666208d99ca37f6d*7C5cdc5b43d7be4caa8173729e3b0a62d9*7C1*7C0*7C637713048300132034*7CUnknown*7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0*3D*7C1000&sdata=EaEp5zLZ7HkDhsBHmP3x3ObPl8j14B8*2BFcOkkNEWZ9w*3D&reserved=0__;JSUlJSUlJSUlJSUlJSUlJSUlJSUlJSUl!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7UzMnNL04$>

Sincerely,
Tsvi Achler

On Wed, Oct 27, 2021 at 2:24 AM Schmidhuber Juergen <juergen at idsia.ch<mailto:juergen at idsia.ch>> wrote:
Hi, fellow artificial neural network enthusiasts!

The connectionists mailing list is perhaps the oldest mailing list on ANNs, and many neural net pioneers are still subscribed to it. I am hoping that some of them - as well as their contemporaries - might be able to provide additional valuable insights into the history of the field.

Following the great success of massive open online peer review (MOOR) for my 2015 survey of deep learning (now the most cited article ever published in the journal Neural Networks), I've decided to put forward another piece for MOOR. I want to thank the many experts who have already provided me with comments on it. Please send additional relevant references and suggestions for improvements for the following draft directly to me at juergen at idsia.ch<mailto:juergen at idsia.ch>:

https://people.idsia.ch/~juergen/scientific-integrity-turing-award-deep-learning.html<https://urldefense.com/v3/__https:/nam12.safelinks.protection.outlook.com/?url=https*3A*2F*2Fpeople.idsia.ch*2F*juergen*2Fscientific-integrity-turing-award-deep-learning.html&data=04*7C01*7Clevine*40uta.edu*7Cb1a267e3b6a64ada666208d99ca37f6d*7C5cdc5b43d7be4caa8173729e3b0a62d9*7C1*7C0*7C637713048300142030*7CUnknown*7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0*3D*7C1000&sdata=mW3lH7SqKg4EuJfDwKcC2VhwEloC3ndh6kI5gfQ2Ofw*3D&reserved=0__;JSUlJX4lJSUlJSUlJSUlJSUlJQ!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7UNznV_Qo$>

The above is a point-for-point critique of factual errors in ACM's justification of the ACM A. M. Turing Award for deep learning and a critique of the Turing Lecture published by ACM in July 2021. This work can also be seen as a short history of deep learning, at least as far as ACM's errors and the Turing Lecture are concerned.

I know that some view this as a controversial topic. However, it is the very nature of science to resolve controversies through facts. Credit assignment is as core to scientific history as it is to machine learning. My aim is to ensure that the true history of our field is preserved for posterity.

Thank you all in advance for your help!

Jürgen Schmidhuber

--
Gary Cottrell 858-534-6640 FAX: 858-534-7029
Computer Science and Engineering 0404
IF USING FEDEX INCLUDE THE FOLLOWING LINE:
CSE Building, Room 4130
University of California San Diego                                      -
9500 Gilman Drive # 0404
La Jolla, Ca. 92093-0404
Email: gary at ucsd.edu<mailto:gary at ucsd.edu>
Home page: http://www-cse.ucsd.edu/~gary/<https://urldefense.com/v3/__http:/www-cse.ucsd.edu/*gary/__;fg!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7U-G68mLE$>
Schedule: http://tinyurl.com/b7gxpwo<https://urldefense.com/v3/__http:/tinyurl.com/b7gxpwo__;!!IKRxdwAv5BmarQ!P43fgF97h1EkMmUyqwIyGb3BiM6QvDDIayyZy_zt_11O7NVqPb6YiU7UcMz40H8$>

Listen carefully,
Neither the Vedas
Nor the Qur'an
Will teach you this:
Put the bit in its mouth,
The saddle on its back,
Your foot in the stirrup,
And ride your wild runaway mind
All the way to heaven.

-- Kabir
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20211107/f740a23b/attachment.html>