“Epiplexity (epistemic complexity) is a formal information-theoretic framework introduced by researchers at NYU and CMU to measure learnable structure for computationally bounded systems like neural networks. It fundamentally challenges classical Shannon entropy by accounting for observer limitations.”
“Core Framework Paradoxes Solved
- Deterministic Functions: Explains how deterministic simulations generate new, learnable structural information for bounded observers.
- Sequence Dependence: Accounts for data order and training curricula, which classical symmetric metrics ignore.
- Inductive Biases: Focuses on the structural shortcuts internalized during training rather than pure distribution matching.”
Primary Resources
- Research Paper: Read the foundational paper on arXiv: From Entropy to Epiplexity
- Source Code: Explore the PyTorch/JAX implementation on GitHub: epiplexity
- Video Lecture: Watch Andrew Gordon Wilson’s presentation on YouTube: Epiplexity for OOD Generalization
- Launch Discussion: Review the community reception on LinkedIn: Epiplexity Announcement Thread
