Song Han

Compressing and accelerating neural networks for deployment.

Han studies how to make neural networks efficient enough to deploy. These papers connect pruning and quantization with memory-efficient language-model inference and efficient generative models, considering both algorithms and the hardware executing them.

Selected work

10 papers

2015 · Efficient AI

Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding

Deep Compression combines pruning, weight quantization, and coding to reduce the storage requirements of neural networks.

Paper & context

2023 · Quantization

AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration

AWQ uses activation information to choose weight quantization strategies that preserve model performance under compression.

Paper & context

2023 · LLMs

Efficient Streaming Language Models with Attention Sinks

Attention Sinks studies why a few initial tokens can preserve streaming language-model performance over long sequences, enabling efficient generation with limited memory.

Paper & context

2022 · Efficient AI

SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models

SmoothQuant makes weight-and-activation quantization easier by redistributing the effect of activation outliers.

Paper & context

2019 · Efficient AI

Once-for-All: Train One Network and Specialize it for Efficient Deployment

Once-for-All trains a network that contains many usable subnetworks for different deployment constraints.

Paper & context

2026 · Efficient AI