Connectionists: LeCun on Marcus

[Danny Silver]

Dear Dr. Marcus .. Thank you for forwarding the link of Dr. LeCun’s article of June 16 to the connectionists list.
I read this with great interest and found it falls very much in line with my thoughts over the last few years.
My sense is that this article will prove pivotal in the debate over the role that symbols have played in human intelligence and their place in AI systems of the future.

For over 30 years, many of us working in machine learning and more specifically neural networks have felt that symbols play more of a peripheral role in intelligence.  And of late … because of work in deep networks, autoencoders, lifelong machine learning, and transformer networks that develop structured vectorial embeddings of concepts (particularly from unsupervised examples) and without the a prioi need for symbols … new theories have arisen of how vectorial representation and symbolic representation intertwine.

My perspective on symbols has become roughly as follows:  Symbols are external communication tools used between intelligent agents that allows knowledge to be transferred in a more efficient and effective manner than having to experience the world directly.  They are also used internally within an agent through a form of self-communication to help formulate and justify decision making. Therefore, symbols are critical to intelligence NOT because they are the building blocks of thought, but because they are characterizations of thought that act as constraints on learning about the world.

Please let me explain.   The ability of one agent to inform another agent that the object they are seeing “can be eaten” or “can eat them”, is a powerful evolutionary pressure.   So, one can see why the basic use of symbols (crows cry to each other, wolves howl) have evolved to be used by so many species.  Humans and several other species have turned it up a notch or two.  So, I agree with Dr. LeCun when he writes “the machine can learn to manipulate symbols in the world, despite not having hand-crafted symbols and symbolic manipulation rules built in” and “This treats symbols and symbolic manipulations as primarily cultural inventions, dependent less on hard wiring in the brain and more on the increasing sophistication of our social lives.”  Similarly, much of human reasoning (perhaps the most important of human decision making, such as “should I start that business”, or “should I marry that man”) is not built on a symbol manipulating mechanism.

However … and this is where my perspective may differ from or add to that of Dr. LeCun … symbols have come to play a very important role in formulating and justifying human decision making to the point where they are critical to the process of intelligent behaviour.  And this is because something very special happens when you put together the following two components or subsystems:  one that learns to sense the world and act appropriately and another that learns to sense its own internal representations and related them to shared symbols.   I suspect these components can be setup to beneficially constrain each other.

The ability to sense the world and react to it appropriately (so as to survive) can be done without being able to externally or internally manipulate symbols (at least in an complex manner).  Many simple creatures do this.  However, if an agent has the capacity to recognize internal vectorial concepts (neuron activity),  label them with symbols, and use those symbols to communicate, then perhaps some marvelous things happen (the agent not only survives, it thrives):

  *   First and foremost, symbols provide the means by which knowledge within the nervous system of one agent can be used to inform the nervous system of another agent, without that second agent expending energy in the real world or risking its life or limb.  This may be one reason why we are the dominate species on the planet.
  *   Second, and perhaps of greater importance, the ability to relate internal neural activity with external symbols and vice versa, provides a mechanism by which an agent can consciously work to form and justify a decision.  This form of internal agent “self-communication” using the same symbols meant for agent-to-agent communication, has become key to human intelligence because it places an additional constraint on learning.  What we learn is forced to fit into the “lexicon” of what we recognize as symbols.


What I am suggesting is, from an evolutionary perspective:
(1) the need to more effectively and efficiently learn about the world and act appropriately led to the ability to recognize internal neural states and related those with shared symbols, and
(2) the development of shared symbols and language provided an additional and beneficial constraint on learning about the world.

Respectfully,
Danny Silver

==========================
Daniel L. Silver
Professor, Jodrey School of Computer Science
Director, Acadia Institute for Data Analytics
Acadia University,
Office 314, Carnegie Hall,
Wolfville, Nova Scotia Canada B4P 2R6
Cell: (902) 679-9315

acadiau.ca<http://www2.acadiau.ca/index.php>
Facebook<https://www.facebook.com/acadiauniversity>  Twitter<https://twitter.com/acadiau>  YouTube<https://www.youtube.com/user/AcadiaWebmaster>  LinkedIn<https://www.linkedin.com/company/acadia-university?trk=biz-companies-cym>  Flickr<https://www.flickr.com/photos/acadiauniversity/albums>
[id:image001.png at 01D366AF.7F868A70]




From: Connectionists <connectionists-bounces at mailman.srv.cs.cmu.edu> on behalf of Gary Marcus <gary.marcus at nyu.edu>
Date: Friday, June 17, 2022 at 3:34 AM
To: Connectionists List <connectionists at cs.cmu.edu>
Subject: Connectionists: LeCun on Marcus
CAUTION: This email comes from outside Acadia. Verify the sender and use caution with any requests, links or attachments.
I’ll probably write a bit of reply later, but this is an excellent new essay by Yann LeCun, quite relevant to many recent discussions here:

https://www.noemamag.com/what-ai-can-tell-us-about-intelligence
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20220617/97a891df/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: image001.png
Type: image/png
Size: 7867 bytes
Desc: image001.png
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20220617/97a891df/attachment.png>