Taxonomy

A hashing model takes an input data-point e.g. an image or document, and outputs a sequence of bits (hashcode) representing that data-point. Hashing models can be broadly categorised into two different categories: quantisation and projection. The projection models focus on learning a low-dimensional transformation of the input data in a way that encourages related data-points to be closer together in the new space. In contrast, the quantisation models seek to convert those projections into binary by using a thresholding mechanism. The projection branch can be further divided into data-independent, data-dependent (unsupervised) and data-dependent (supervised) depending on whether the projections are influenced by the distribution of the data or available class-labels.