Interpreting Emergent Communication
summary
Describes an emergent-communication pipeline that inserts a codebook discrete module between sender and receiver, uses bipartite matching to map symbols to dataset concepts, and achieves strong zero-shot compositional generalization on synthetic tasks (digits/shapes) by concatenating learned vectors. Proves that optimal reconstruction games force semantic consistency (similar inputs map to similar messages) while discrimination games can encode arbitrary distinctions, even noise. Demonstrates unsupervised translation on MS COCO requiring hundreds of thousands of messages and yielding coarse theme-level English output.
AI-written summary · about
chapters (7)
- 0:00Emergent communication intro
- 4:31Understanding emergent language
- 10:16Evaluating compositionality
- 17:59Meaningful communication in agents
- 26:53Theoretical limits of communication
- 32:26Unsupervised translation
- 44:34Methods for animal communication
related talks
52mGeneralization in the representations and computations of frontier language models.
01h 06mMixed-modal Language Modeling: Chameleon, Transfusion, and Mixture of Transformers
37mShort Industry Talks pt.2
03mContextual understanding using Generative Artificial Intelligence
31mText and Code Embeddings
58mMulti-Agent Reinforcement Learning Towards Zero-Shot Communication
43mDiscussion
01h 22mBeyond Decoding: Meta-Generation Algorithms for Large Language Models