AI Paper Atlas

Research map / Architectures and efficiency / Quantization

Matrix Compression via Randomized Low Rank and Low Precision Factorization

Rajarshi Saha, Varun Srivastava, Mert Pilanci

NeurIPS 2023 · Architectures and efficiency › Quantization · approach: Post-training weight quantization · 5 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
replace naive quantization with LPLR at 1 bitembedding classifier · CIFAR-10 embeddingsaccuracy (%): 11 -> 92Improved
replace naive quantization with LPLR at 1 bitembedding classifier · CIFAR-100 embeddingsaccuracy (%): 1 -> 79Improved
replace naive quantization with LPLR at 1 bitembedding classifier · IMDB embeddingsaccuracy (%): 50 -> 73Improved
replace naive quantization with LPLR at 1 bitembedding classifier · Emotion embeddingsaccuracy (%): 29 -> 41Improved
reduce LPLR factor bits from B=32 to B=8 (larger sketch) at B_nq=1 (ablation)LPLR · Shepp-Logan phantomrelative Frobenius norm error: 0.610 -> 0.340Improved
LPLR instead of naive quantization at B_nq=1, B=8LPLR · Shepp-Logan phantomrelative Frobenius norm error: 0.532 -> 0.340Improved
LPLR instead of naive quantization at B_nq=1, B=32LPLR · Shepp-Logan phantomrelative Frobenius norm error: 0.532 -> 0.610Hurt
LPLR instead of naive quantization at B_nq=2, B=32LPLR · Shepp-Logan phantomrelative Frobenius norm error: 0.312 -> 0.447Hurt
use SVD-based left factor (LPLR-SVD) instead of random sketch (ablation)LPLR · Shepp-Logan phantom, B=32, m=15relative Frobenius norm error: 0.610 -> 0.506Improved
use SVD-based left factor (LPLR-SVD) instead of random sketch (ablation)LPLR · IMDB embeddings, B_nq=1relative Frobenius norm error: 0.313 -> 0.241Improved
LPLR instead of direct-SVD quantizationLPLR · CIFAR-100 embeddings, B_nq=1accuracy (%): 82 -> 79Hurt
LPLR instead of direct-SVD quantizationLPLR · IMDB embeddings, B_nq=1relative Frobenius norm error: 0.229 -> 0.313Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy