Toward Efficient Machine Learning Systems with Sampling and Compression.

Abstract

The rapid growth of machine learning - in terms of model size, dataset volume, and task complexity - has created significant computational and memory efficiency challenges. This thesis addresses these challenges by developing sampling and compression techniques across a variety of machine learning applications. Specifically, we introduce sampling strategies for efficient graph neural networks (CacheSample), accelerated 3D image rendering (FastSR-NeRF), and optimized retrieval-augmented generation systems (TeleRAG). We further propose novel compression methods targeting convolutional neural networks (SPIN), large language models (Atom), and their key-value caches (Palu). Collectively, these techniques substantially reduce computational and memory requirements while preserving model accuracy, facilitating the scalability and accessibility of machine learning systems. Finally, I present my vision for future efficiency innovations to ensure continued scalability and robustness as machine learning models continue to grow in complexity.