Grounding Doctrine

From the AI Research Compass wiki

Grounding Doctrine: Symbols mean something only when connected to perception and action; text-only learning leaves meaning ungrounded.

Beliefs

Grounding Doctrine starts from the symbol grounding problem: a system that only shuffles symbols defined by other symbols is like someone learning a language from a dictionary written in that same language. Somewhere the chain must bottom out in perception and action, in pixels, sounds, forces and outcomes. A model that has read about ‘heavy’ a million times but never lifted anything knows something, but not all of what the word means.

It sits far toward the World edge, in the upper half: perception is where meaning is anchored, though language remains a welcome partner rather than an enemy. The doctrine accepts grounding through many routes, from image–text pairs to referring expressions to robot interaction, so it is pluralist about architecture. Its ultra-revolutionary shade reflects a long-standing challenge to the text-first mainstream that has recently become an engineering programme, as multimodal models make grounding measurable through tasks like visual question answering and referring-expression localization.

Arguments

  • Text-only models make characteristic errors on spatial, physical and perceptual questions that grounded multimodal models handle better.
  • Grounded tasks such as referring-expression localization give operational tests of whether a word is actually connected to the world.
  • Language acquisition in humans happens alongside perception and interaction, suggesting grounding is the natural route to meaning.

Critiques

  • Distributional semanticists note that text-only models already capture much perceptual knowledge indirectly, because people write about what they see.
  • It is unclear how much grounding is enough, so the doctrine can retreat to ever-stronger requirements whenever text-only models succeed.
  • Image–text pairs ground language in other people’s photos, which is a thin and biased version of experience.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ +1.00, y ≈ +0.61, paradigm ≈ +0.74.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data