Reproduce “Morphing-based Compression for Data-centric ML Pipelines” 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

Sebastian Baunsgaard, Matthias Boehm

BWARE is a compression framework that accelerates data-centric machine learning pipelines by exploiting data redundancy throughout feature engineering and model training. The framework leverages lightweight frame compression and a novel morphing technique to enable workload-aware optimization of compressed representations, leading to significant reductions in memory usage and faster execution times compared to uncompressed workflows.

Key insights

  • Real-world machine learning datasets exhibit significant value-based, structural, and correlation-based redundancy that can be exploited for compression.
  • A lightweight morphing technique allows for workload-aware optimization of compressed data representations without requiring full decompression, adapting to different ML operations.
  • Integrating compression directly into the data preparation and feature engineering stages of ML pipelines, rather than just model training, yields substantial end-to-end performance gains.

Summary by alphaXiv.

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