SimSiam Naming Game: A Unified Approach for Emergent Communication and Representation Learning
Researchers have introduced the SimSiam Naming Game (SSNG), a novel framework for emergent communication that bypasses the sample-inefficiency of previous methods like the Metropolis-Hastings Naming Game (MHNG). SSNG utilizes a self-supervised representation alignment objective between autonomous agents, enabling end-to-end gradient-based optimization through a Gumbel-Softmax relaxation for discrete symbolic messages. Experiments on CIFAR-10 and ImageNet-100 datasets demonstrate that SSNG's emergent messages achieve superior classification accuracy compared to existing referential and reconstruction games, as well as MHNG, highlighting its effectiveness in feedback-free emergent communication for multi-agent systems. AI