Jethro's Braindump

And the Bit Goes Down: Revisiting the Quantization of Neural Networks

Model Compression
(Stock et al., n.d.)

This method minimizes the loss reconstruction error for in-domain inputs, and does not require any labelled data.

This method exploits the high correlation in the convolutions in ResNet-like architectures by the use of product quantization (PQ). The approach here focuses on reconstructing the activations, and not the weights. This results in better in-domain reconstruction, and does not require any supervision.

Vector Quantization (VQ) and Product Quantization (PQ) decompose the high-dimensional space into a cartesian product of subspaces that are quantized separately. These are typically studied under the context of nearest neighbour search.


Stock, Pierre, Armand Joulin, Rémi Gribonval, Benjamin Graham, and Hervé Jégou. n.d. “And the Bit Goes down: Revisiting the Quantization of Neural Networks.”