11 / Inductive learning

Loom

Teach a small learner with examples and watch competing rules disappear.

Interactive arithmetic illustration · saved grammar benchmark

11 / Examples into programs

Teach it a rule.

One example can fit many rules. Add an example to narrow the candidates, then see what they predict on unseen inputs.

Teaching examples

    Surviving hypotheses

    Where the candidates disagree

    Darker squares: agreement. Orange squares: more disagreement. Grid inputs run from −3 to 3.

    A small browser illustration of Loom’s arithmetic grammar. Teaching stays in this tab. The full local learner also handles lists, text, retrieval, and persistent memory.

    The saved numeric benchmark

    The saved benchmark passed 60 of 60 held-out cases across ten families drawn from the implemented grammar, using four examples per family.

    The project

    A learner with inspectable hypotheses

    Loom searches a typed program grammar from input/output examples, retains competing explanations, and proposes examples that distinguish them. Its local workbench adds a randomly initialized contrastive encoder, experience replay, retrieval, and persistent SQLite memory.

    Observations & lessons

    What a single example leaves open

    From 1 + 1 = 2, a constant 2 and the rule x + y both fit. Additional examples distinguish them. The full learner uses separate held-out tests before execution; the saved numeric benchmark passes all 60 grammar-specific cases.