Reproduce “Emergent Communication: Generalization and Overfitting in Lewis Games” on OpenResearch

About this paper

Mathieu Rita, Corentin Tallec, Paul Michel, Jean-Bastien Grill, Olivier Pietquin, Emmanuel Dupoux, Florian Strub

An analytical framework decomposes the Lewis game objective into information and co-adaptation losses, uncovering that overfitting in co-adaptation is a primary barrier to robust emergent communication. Managing this specific overfitting enables artificial agents to develop more generalizable and compositionally structured languages, aligning with human linguistic expectations.

Key insights

• Standard continuous training regimes lead to severe co-adaptation overfitting, where agents specialize in interpreting each other on training data but fail to generalize.
• There is an optimal "sweet spot" for listener convergence: too little co-adaptation pressure leads to poor languages, but excessive co-adaptation leads to detrimental overfitting.
• Controlling co-adaptation overfitting, particularly by managing the listener's learning, demonstrably improves both the generalization and compositionality of emergent languages.

Summary by alphaXiv.