Reproduce “Memory Efficient Tabular Foundation Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Shuting Luo, Monika Mikhail Kanaan, Cameron Gordon, Anna Leontjeva, Simon Lucey

This paper investigates the practical deployment of Tabular Foundation Models (TFMs) by applying post-training quantization techniques to reduce their memory footprint. The research shows that TFMs can achieve up to a 7.6x reduction in deployment memory while retaining their predictive performance advantage over classical machine learning methods.

Key insights

  • Post-training quantization can dramatically reduce the deployment memory footprint of Tabular Foundation Models (TFMs) without meaningfully degrading their predictive performance.
  • K-Means INT4 quantization emerged as the most efficient method, achieving the highest compression ratio with the least accuracy loss among the tested techniques.
  • The optimal bit-width for quantization was found to be 4-bit (INT4), providing over 99% baseline accuracy recovery with substantial compression.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.