• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
  • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Tom McCoy Discusses TPR-Augmented Transformers

    Tom McCoy references a paper on augmenting Transformers with TPR structure.

    TM
    3 Sources, 29d ago, first seen 29d ago

    TLDR

    Tom McCoy, assistant professor at Yale, replied to questions on X about work combining Transformers and tensor product representations. He pointed to an overview paper on neurocompositional computing that describes how adding TPR structure to a standard Transformer yields benefits. In a second reply McCoy noted that EnyanZhang will soon release work connecting TPRs to representational analyses in mechanistic interpretability, while observing that further analysis of mechanisms remains needed.

    Combined views

    284

    3 Sources, first seen 29d ago

    Combined views

    284

    3 Sources, first seen 29d ago

    5 likes
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    5 likes
    2 comments

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    2 comments

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    3 Sources

    @RTomMcCoy@raphaelmilliere @tallinzen @paul_smolensky The other type of evidence I could imagine being relevant is: If a model implements TPRs, I would expect it to show better compositional generalization than if it just approximates them. But I don't know if the existing empirical data points strongly in one way or the other

    3 Sources

    @RTomMcCoy@raphaelmilliere @tallinzen @paul_smolensky The other type of evidence I could imagine being relevant is: If a model implements TPRs, I would expect it to show better compositional generalization than if it just approximates them. But I don't know if the existing empirical data points strongly in one way or the other