Deep Emergent Communication: How It Started and How It’s Going
summary
Presents concrete experiments where neural agents play a communication game (sender generates discrete messages, receiver points to the target) using transformers and a CLIP retriever; agents learn to coordinate on unseen images but often exploit low‑level cues and even succeed on random noise. Adding data augmentation forces semantically coherent categories; imposing a message cost and impatient receivers produces Zipf‑like length–frequency patterns. Pretrained image‑conditioned LMs produce captions that help neural retrievers but not humans, revealing semantic drift and mixed progress toward human‑like language.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 7:59Communication game setup
- 13:10Key results in research
- 24:23Language properties and limits
- 37:55Deep communication with LLMs
- 46:24Future directions
- 50:44Closing remarks and Q&A
related talks
13mInside Multimodal Models
26mMaximizing Multimodal: Exploring the search frontier of text-to-image models
36mNeural Networks
01h 16mGeneralization in natural language processing
42mDALL·E and CLIP from OpenAI: a new era in machine learning
53mGenerative AI: Overview and use cases
34mNeural networks: insane in the membrane
40mPanel discussion "NLP"