AI/MLsrc: curated AI glossary

Quantization

Quantization is like taking a high-resolution photo and saving it as a slightly smaller file; you lose a tiny bit of detail, but the file takes up much less space and loads faster.

Quantization reduces the memory footprint and computational requirements of a model by converting high-precision weight and activation values into lower-bit representations, such as moving from FP16 to INT8, which accelerates inference at the cost of slight precision degradation.

Reducing the numerical precision of a model's weights and/or activations (e.g. 16-bit to 4-bit) to cut memory and speed inference, usually with minor accuracy loss.

evolution

2012 · history
AlexNet and Early Fixed-Point Research
Early research into fixed-point arithmetic for neural networks began to emerge as a way to reduce the computational burden of deep learning models.
2016 · history
Deep Compression
Song Han et al. introduced 'Deep Compression,' a seminal paper demonstrating that pruning, trained quantization, and Huffman coding could reduce model size by 35x-49x without accuracy loss.
2017 · history
Quantization-Aware Training (QAT)
Google researchers popularized QAT, allowing models to learn to compensate for precision loss during the training process rather than just post-training.
2022 · history
LLM Quantization (bitsandbytes)
The release of 8-bit quantization techniques for Large Language Models enabled massive models to run on consumer-grade hardware, significantly democratizing LLM access.
2023 · history
4-bit Quantization (QLoRA)
The introduction of QLoRA allowed for the fine-tuning of massive models on single GPUs by using 4-bit NormalFloat quantization, setting a new standard for efficient model adaptation.

← all terms