Rotation Trick

Restructuring Vector Quantization with the Rotation Trick

Christopher Fifty et al. · ICLR 2025 Oral

Read paper · arXiv:2410.06424

Use a frozen rotation and scale to construct vector-quantization surrogate gradients.

What you will implement

Rotation-based gradient transformation with quantized forward values.

• 41. Rotation Trick Quantization Gradients

  Keep the quantized forward value while transporting encoder gradients through a frozen scaled rotation.

  Given the selected code, implement the gradient surrogate; codebook learning and a full VQ-VAE are outside this module.