Binary Quantisation Models

Quantisation models Two main categories of quantisation have been proposed for nearest neighbour search: scalar and vector quantisation, which are differentiated by whether the input and output of the quantisation is a scalar or a vector quantity. This page lists models for scalar quantisation. Scalar quantisation is frequently applied to quantise the real-values (projections) resulting from the dot product of the feature vector of each data-point onto the normal vectors to a set of hyperplanes partitioning the feature space. Each dot product yields a scalar value which is then subsequently quantised into binary (0/1) by thresholding. The resulting bits are concatenated to form the hashcode for a data-point.

Paper	Codebook	Optimisation	Learning Type	#Thresholds
, .	Binary	Backpropagation	Supervised	N/A
, .	Binary	Backpropagation	Supervised	N/A
, .	Natural Binary Code (NBC)	K-Means	Unsupervised	3+
, .	Binary	Squared Error Minimisation	Unsupervised	2
Yunqiang Li, Wenjie Pei, Yufei Zha, Jan van Gemert, 2019.Push For Quantization: Deep Fisher Hashing	Any	Backpropagation	Supervised	N/A
, .	Any	Stochastic Search	Semi-Supervised	1+
, .	Any	Stochastic Search + Binary Integer Linear Program	Semi-Supervised	Variable
, .	Natural Binary Code (NBC)	Maximum Margin + Stochastic Search	Semi-Supervised	Variable
, .	Any	Greedy Strategy	Unsupervised	Variable
, .	Binary	Squared Error Minimisation	Unsupervised	2
, .	Binary	K-Means	Unsupervised	N/A
, .	Natural Binary Code (NBC)	Dynamic Programming	Unsupervised	Variable