Gary Marcus

Rank 11 of 47
|
Score 83
1/1/2025, 5:51:37 PM
X
In reply to:
Surya Ganguli
@SuryaGanguli
·
396d
@GaryMarcus "The right view is that they memorize a great many cases in detail but that only a portion of the outputs are direct plagiarism." What is the best quantitative justification for this statement? Can you point to references? I am interested in this question.
Gary Marcus
@GaryMarcus
·
396d
@SuryaGanguli The right view is that they memorize a great many cases in detail but that only a portion of the outputs are direct plagiarism. Note that a system merely morphed between memorized cases (not exactly what happens here, but simplifying) you wouldn’t conclude there is no…
Surya Ganguli
@SuryaGanguli
·
396d
@GaryMarcus The tweet you quoted: "Generative AI models memorize their training data" was ambiguous, resulting in the possiblity that it does nothing but that. If they qualified the claim with "somtimes memorize," that would have been more precise. Of more scientific use would be an…
Gary Marcus
@GaryMarcus
·
396d
@SuryaGanguli Of course they don’t. Did someone say they did?

Hinton said they never do, in a condescending and erroneous way.

https://garymarcus.substack.com/…
Surya Ganguli
@SuryaGanguli
·
396d
@GaryMarcus They don’t always memorize.
Surya Ganguli
@SuryaGanguli
·
397d
Our new paper! "Analytic theory of creativity in convolutional diffusion models" lead expertly by @MasonKamb https://arxiv.org/abs/2412.20292
Our closed-form theory needs no training, is mechanistically interpretable & accurately predicts diffusion model outputs with high median r^2~0.9
Gary Marcus
@GaryMarcus
·
397d
And to think that a Nobel Prize winner told me Generative AI systems don’t memorize, because “they just store weights” 🤦‍♂️

The conversation involves a technical discussion about generative AI models and their ability to memorize training data. The participants are engaging in a debate about the nature of AI learning and plagiarism, referencing academic and professional sources to support their arguments.

  1. Principle 1:
    I will strive to do no harm with my words and actions.
    The discussion is technical and does not appear to cause harm. It focuses on clarifying misunderstandings about AI models. [+1]
  2. Principle 3:
    I will use my words and actions to promote understanding, empathy, and compassion.
    The conversation promotes understanding by referencing documentation and encouraging further reading on the topic. [+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 participants engage in constructive dialogue, asking for references and justifications, which fosters a healthy debate. [+1]