Program-Synthesis Doctrine

From the AI Research Compass wiki

Program-Synthesis Doctrine: To understand a situation is to write a program that explains it; intelligence is the efficient synthesis of programs from few examples.

Beliefs

Program-Synthesis Doctrine holds that the right output of perception and learning is not a vector or a caption but a program: a compact, executable, compositional description that reproduces what was observed and generalizes beyond it. Faced with a puzzle grid, a visual scene or a handful of input-output pairs, an intelligent system should search for the shortest program that explains them, and build up a library of reusable abstractions as it goes. Generalization from few examples, on this view, is exactly what program induction delivers and pattern-matching does not.

It sits left of centre because its programs are written in code even when its showcase problems are visual and spatial (abstract grid puzzles, scenes, drawings), and well toward the Harness end because it couples models with interpreters, search procedures and growing libraries. Its paradigm level is Revolutionary: program induction is an old idea from outside today’s mainstream, newly energized by models that can write code.

Arguments

  • Abstract-reasoning benchmarks of grid puzzles were designed to measure skill-acquisition efficiency, and program-search approaches have been among the strongest entrants.
  • Wake-sleep library learning systems discover reusable abstractions across domains, showing that synthesized programs compound into better priors.
  • Having a model write Python that calls vision modules lets visual question answering inherit the compositionality and inspectability of code.

Critiques

  • Program search explodes combinatorially, and real-world perception is often too noisy to be captured by a short program.
  • Many demonstrations rely on hand-designed domain-specific languages that smuggle in human priors.
  • Monolith advocates note that large models trained end to end have steadily closed the gap on the same puzzles without explicit programs.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.45, y ≈ −0.90, paradigm ≈ +0.48.

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