| Название исследуемой задачи: | Epiplexity on one GPU: estimators for small networks and when they agree |
|---|---|
| Тип научной работы: | CoIS, BMM |
| Авторы: | Алексей Кравацкий, Марк Иконников, Даниил Казачков, Данила Черноусов |
| Научный руководитель: | уточняется |
| Научный консультант(при наличии): | уточняется |
Epiplexity (Finzi, Qiu et al., 2026) measures the structure
that a computationally bounded learner can extract from data. This project estimates epiplexity
with small models (MLPs and convolutional networks) on small datasets on a single GPU. The
epimeter library gives one interface that returns the model part and the data part of a code
length in bits for the prequential, requential, variational and Laplace codes, the classical
criteria, EDL and the reservoir score. With it we estimate epiplexity on elementary cellular
automata, label noise, permuted pixels and a planted teacher–student task, measure the spread
over seeds under a fixed protocol, and find when Bayesian model parts agree with epiplexity.
- A python package epimeter with all implementation here.
- Runnable examples wired through the registries here. The Colab demo notebook will be added to the same folder.
src/epimeter/ the package: estimators, models, samplers, training, eval
tests/ pytest
code/ experiment code and demo
doc/ Sphinx documentation
paper/ arXiv-style manuscript
figures/ figures for the paper
slides/ presentations: project/ (project plan), talk/ (talk on the epiplexity paper)
blogpost/ blog post and its preview script
notes/ literature review (pdf + tex + refs)
research_notes/ source notes behind the literature review
plan/ plan, roles, protocol, calendar
description.md architecture and public API of the package
Requires Python 3.10+ and uv. On a network that needs system TLS
certificates, add --system-certs to the uv commands.
uv sync --extra dev
.venv/bin/pytest tests/
.venv/bin/ruff check src/ tests/ code/