Self-supervised training of hierarchical vector quantisers

The author has previously developed a hierarchical vector quantisation (VQ) model which successfully applied to time series and image compression respectively. The paper derives an extension to this model, in which the author backpropagates signals from higher to lower layers of the hierarchy to self-supervise the training of the VQ. He reviews the basic properties of his VQ model and its relationship to neural network methods. He extends the model to an ensemble of VQs, and derives its properties in the limit of a large codebook size (i.e. the continuum limit). Finally, he demonstrates how self-supervision emerges naturally in this type of model.<
>