Bayesian methods are used to estimate intrinsic dimensionality (ID) and to perform dimensionality reduction for a large complex dataset. Using 130,000 images over 100 categories, we developed a process to reduce the dimensionality to a very small size while preserving the ability to classify the images. The novelty of our approach is two fold, 1) 2NN estimation of target dimensionality is now used as a prior distribution, and 2) the overall mapping is generated by a Bayesian neural network which then selects an appropriate intrinsic dimension based on the prior, variational inference, and multidimensional scaling.
Read
