Reproduce “Testing GLOM's ability to infer wholes from ambiguous parts” on OpenResearch

About this paper

Laura Culp, Sara Sabour, Geoffrey E. Hinton

The eGLOM system tests whether parts with ambiguous identities can agree on shared objects. It uses symbolic ellipse inputs: each part proposes a whole, and attention-weighted agreement combines similar proposals. In supervised 2-from-2 tests, eGLOM’s Whole MSE was 9.6 × 10⁻⁶, versus 6.1 × 10⁻⁴ for an MLP autoencoder; this evidence is limited to a simplified, labeled setting, not pixel vision.

Key insights

• The GLOM architecture's principle of multi-modal part predictions combined with attention-weighted averaging effectively resolves ambiguity and forms dynamic 'islands of similarity' for hierarchical object parsing.
• eGLOM's recurrent updates facilitate the correction of highly perturbed input parts by leveraging the global object context established through consistent whole predictions.
• The architectural principles demonstrated strong generalization to out-of-distribution transformations and exhibited high parameter efficiency compared to a traditional autoencoder baseline.

Summary by alphaXiv.