Reproduce “Permutation-Invariant Set Autoencoders with Fixed-Size Embeddings for Multi-Agent Learning” on OpenResearch

About this paper

Ryan Kortvelesy, Steven Morad, Amanda Prorok

The Permutation-Invariant Set Autoencoder (PISA) offers a method to encode variable-sized sets into fixed-size latent embeddings and reconstruct them, designed for multi-agent systems. The architecture achieves near-perfect reconstruction (e.g., 0.9999 correlation for R^96 latent size) and a similarity-preserving latent space, facilitating unsupervised communication in multi-agent sensor fusion tasks.

Key insights

• The deterministic key-value pairing mechanism allows for a direct correspondence between input and reconstructed elements, simplifying loss calculation and enhancing training efficiency by avoiding complex matching algorithms.
• Achieving permutation-invariance through element-wise key-value products aggregated by summation, combined with cardinality encoding, results in a fixed-size, similarity-preserving latent space robust to input order.
• PISA's architecture enables a novel capability to manipulate latent states by directly inserting or removing elements or combining sets, offering flexibility not typically found in other set encoding methods.

Summary by alphaXiv.