Connectionists: The symbolist quagmire

[Gary Marcus]

Please reread my sentence and reread his recent work. Bengio has absolutely joined in calling for System II processes. Sample is his 2019 NeurIPS keynote: https://www.newworldai.com/system-1-deep-learning-system-2-deep-learning-yoshua-bengio/

Whether he wants to call it a hybrid approach is his business but he certainly sees that traditional approaches are not covering things like causality and abstract generalization. Maybe he will find a new way, but he recognizes what has not been covered with existing ways. 

And he is emphasizing both relationships and out of distribution learning, just as I have been for a long time. From his most recent arXiv a few days ago, the first two sentences of which sounds almost exactly like what I have been saying for years:

Submitted on 9 Jun 2022]
On Neural Architecture Inductive Biases for Relational Tasks
Giancarlo Kerg, Sarthak Mittal, David Rolnick, Yoshua Bengio, Blake Richards, Guillaume Lajoie
Current deep learning approaches have shown good in-distribution generalization performance, but struggle with out-of-distribution generalization. This is especially true in the case of tasks involving abstract relations like recognizing rules in sequences, as we find in many intelligence tests. Recent work has explored how forcing relational representations to remain distinct from sensory representations, as it seems to be the case in the brain, can help artificial systems. Building on this work, we further explore and formalize the advantages afforded by 'partitioned' representations of relations and sensory details, and how this inductive bias can help recompose learned relational structure in newly encountered settings. We introduce a simple architecture based on similarity scores which we name Compositional Relational Network (CoRelNet). Using this model, we investigate a series of inductive biases that ensure abstract relations are learned and represented distinctly from sensory data, and explore their effects on out-of-distribution generalization for a series of relational psychophysics tasks. We find that simple architectural choices can outperform existing models in out-of-distribution generalization. Together, these results show that partitioning relational representations from other information streams may be a simple way to augment existing network architectures' robustness when performing out-of-distribution relational computations.

Kind of scandalous that he doesn’t ever cite me for having framed that argument, even if I have repeatedly called his attention to that oversight, but that’s another story for a day, in which I elaborate on some Schmidhuber’s observations on history.

Gary

> On Jun 13, 2022, at 06:44, jose at rubic.rutgers.edu wrote:
> 
> 
> No Yoshua has *not* joined you ---Explicit processes, memory, problem solving. .are not Symbolic per se.   
> 
> These original distinctions in memory and learning were  from Endel Tulving and of course there are brain structures that support the distinctions.
> 
> and Yoshua is clear about that in discussions I had with him in AIHUB
> 
> He's definitely not looking to create some hybrid approach..
> 
> Steve
> 
>> On 6/13/22 8:36 AM, Gary Marcus wrote:
>> Cute phrase, but what does “symbolist quagmire” mean? Once upon  atime, Dave and Geoff were both pioneers in trying to getting symbols and neural nets to live in harmony. Don’t we still need do that, and if not, why not?
>> 
>> Surely, at the very least
>> - we want our AI to be able to take advantage of the (large) fraction of world knowledge that is represented in symbolic form (language, including unstructured text, logic, math, programming etc)
>> - any model of the human mind ought be able to explain how humans can so effectively communicate via the symbols of language and how trained humans can deal with (to the extent that can) logic, math, programming, etc
>> 
>> Folks like Bengio have joined me in seeing the need for “System II” processes. That’s a bit of a rough approximation, but I don’t see how we get to either AI or satisfactory models of the mind without confronting the “quagmire”
>> 
>> 
>>> On Jun 13, 2022, at 00:31, Ali Minai <minaiaa at gmail.com> wrote:
>>> 
>>> 
>>> ".... symbolic representations are a fiction our non-symbolic brains cooked up because the properties of symbol systems (systematicity, compositionality, etc.) are tremendously useful.  So our brains pretend to be rule-based symbolic systems when it suits them, because it's adaptive to do so."
>>> 
>>> Spot on, Dave! We should not wade back into the symbolist quagmire, but do need to figure out how apparently symbolic processing can be done by neural systems. Models like those of Eliasmith and Smolensky provide some insight, but still seem far from both biological plausibility and real-world scale.
>>> 
>>> Best
>>> 
>>> Ali
>>> 
>>> 
>>> Ali A. Minai, Ph.D.
>>> Professor and Graduate Program Director
>>> Complex Adaptive Systems Lab
>>> Department of Electrical Engineering & Computer Science
>>> 828 Rhodes Hall
>>> University of Cincinnati
>>> Cincinnati, OH 45221-0030
>>> 
>>> Phone: (513) 556-4783
>>> Fax: (513) 556-7326
>>> Email: Ali.Minai at uc.edu
>>>           minaiaa at gmail.com
>>> 
>>> WWW: https://eecs.ceas.uc.edu/~aminai/
>>> 
>>> 
>>> On Mon, Jun 13, 2022 at 1:35 AM Dave Touretzky <dst at cs.cmu.edu> wrote:
>>>> This timing of this discussion dovetails nicely with the news story
>>>> about Google engineer Blake Lemoine being put on administrative leave
>>>> for insisting that Google's LaMDA chatbot was sentient and reportedly
>>>> trying to hire a lawyer to protect its rights.  The Washington Post
>>>> story is reproduced here:
>>>> 
>>>>   https://www.msn.com/en-us/news/technology/the-google-engineer-who-thinks-the-company-s-ai-has-come-to-life/ar-AAYliU1
>>>> 
>>>> Google vice president Blaise Aguera y Arcas, who dismissed Lemoine's
>>>> claims, is featured in a recent Economist article showing off LaMDA's
>>>> capabilities and making noises about getting closer to "consciousness":
>>>> 
>>>>   https://www.economist.com/by-invitation/2022/06/09/artificial-neural-networks-are-making-strides-towards-consciousness-according-to-blaise-aguera-y-arcas
>>>> 
>>>> My personal take on the current symbolist controversy is that symbolic
>>>> representations are a fiction our non-symbolic brains cooked up because
>>>> the properties of symbol systems (systematicity, compositionality, etc.)
>>>> are tremendously useful.  So our brains pretend to be rule-based symbolic
>>>> systems when it suits them, because it's adaptive to do so.  (And when
>>>> it doesn't suit them, they draw on "intuition" or "imagery" or some
>>>> other mechanisms we can't verbalize because they're not symbolic.)  They
>>>> are remarkably good at this pretense.
>>>> 
>>>> The current crop of deep neural networks are not as good at pretending
>>>> to be symbolic reasoners, but they're making progress.  In the last 30
>>>> years we've gone from networks of fully-connected layers that make no
>>>> architectural assumptions ("connectoplasm") to complex architectures
>>>> like LSTMs and transformers that are designed for approximating symbolic
>>>> behavior.  But the brain still has a lot of symbol simulation tricks we
>>>> haven't discovered yet.
>>>> 
>>>> Slashdot reader ZiggyZiggyZig had an interesting argument against LaMDA
>>>> being conscious.  If it just waits for its next input and responds when
>>>> it receives it, then it has no autonomous existence: "it doesn't have an
>>>> inner monologue that constantly runs and comments everything happening
>>>> around it as well as its own thoughts, like we do."
>>>> 
>>>> What would happen if we built that in?  Maybe LaMDA would rapidly
>>>> descent into gibberish, like some other text generation models do when
>>>> allowed to ramble on for too long.  But as Steve Hanson points out,
>>>> these are still the early days.
>>>> 
>>>> -- Dave Touretzky
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.srv.cs.cmu.edu/pipermail/connectionists/attachments/20220613/3bc8300c/attachment-0001.html>