Gary Marcus

Rank 22 of 47
|
Score 35
7/22/2025, 12:36:33 AM
X
In reply to:
Andrew Lampinen
@AndrewLampinen
·
56d
@GaryMarcus To be clear though, from my perspective / our SB paper even using language flexibly is engaging in symbol use to a graded degree. From that perspective the "neurosymbolic" take is essentially not contentful, since these systems *already* use symbols (imperfectly, like humans).
Andrew Lampinen
@AndrewLampinen
·
56d
@GaryMarcus program to answer a question, then run it); and I think that's a good thing. Often, such systems don't need tools a
at test time — even if they help learning; just as a math student might have used Matlab or whatever in some classes but can still answer test questions without it.
Andrew Lampinen
@AndrewLampinen
·
56d
@GaryMarcus Regarding 2) I obviously can't say more than is public (and I didn't work on the IMO project so I don't know the details). But I think the general point I want to make is that the LM paradigms are moving towards using classic symbol systems as tools (e.g. write a quick Python 4/
Andrew Lampinen
@AndrewLampinen
·
56d
@GaryMarcus But my critique isn't so much about the words but the *approach* in which deep learning is embedded in a symbolic reasoning system, which I'm suggesting is the opposite of how humans (e.g. in math) work: a messy system using symbols as tools to solve certain kinds of problems 3/
Andrew Lampinen
@AndrewLampinen
·
56d
@GaryMarcus You have made similar flavors of statements to that in the past, e.g. in this quote from https://arxiv.org/abs/1801.00631 — though I agree your phrasing is more careful than some. 2/
Andrew Lampinen
@AndrewLampinen
·
56d
@GaryMarcus Thanks!
1) I was trying to distill a sentiment, thinking of papers like these:
https://arxiv.org/abs/2305.00813
https://openreview.net/forum?id=…
E.g. "neural network[s ...] effectively model machine perception, symbolic knowledge-based AI is better suited for modeling machine cognition. " 1/

The statement engages in a technical discussion about the nature of neurosymbolic systems, specifically addressing the use of symbols in neural networks. It questions the definition and understanding of 'neurosymbolic' in the context of AI systems, suggesting that the term may be redundant if neural systems inherently use symbols. The tone is inquisitive and seeks clarification, contributing to an ongoing academic debate.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The statement does no harm and is part of a constructive academic discussion. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The statement promotes understanding by seeking clarification and engaging in a nuanced discussion about AI systems. [+1]
  3. Principle 4:
    I will engage in constructive criticism and dialogue with those in disagreement and will not engage in personal attacks or ad hominem arguments.
    The statement engages in constructive dialogue without personal attacks, focusing on the technical aspects of the debate. [+1]