Reproduce “Mem-Rec: Memory Efficient Recommendation System using Alternative Representation” 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

Gopi Krishna Jha, Anthony Thomas, Nilesh Jain, Sameh Gobriel, Tajana Rosing, Ravi Iyer

Intel Labs and UCSD researchers developed MEM-REC, a memory-efficient recommendation system utilizing a dual-encoder Bloom filter architecture to manage high-cardinality categorical features. The system achieved up to 2904x memory compression and 3.4x faster embedding lookups while preserving recommendation quality, by enabling embedding tables to fit within CPU caches.

Key insights

  • The dual-encoder Bloom filter architecture successfully mitigates the accuracy degradation typically associated with hash collisions in aggressive embedding compression.
  • MEM-REC's memory footprint scales logarithmically with the alphabet size of categorical features, providing robust scalability for extremely large datasets.
  • By designing embedding tables that fit entirely within the Last-Level Caches (LLCs) of modern server CPUs, the system drastically reduces main memory accesses, improving performance and energy efficiency.

Summary by alphaXiv.

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