Thinner Latent Spaces: Detecting dimension and imposing invariance through autoencoder gradient constraints

Conformal Autoencoders are a neural network architecture that imposes orthogonality conditions between the gradients of latent variables towards achieving disentangled representations of data. In this letter we show that orthogonality relations within the latent layer of the network can be leveraged to infer the intrinsic dimensionality of nonlinear manifold data sets (locally characterized by the dimension of their tangent space), while simultaneously computing encoding and decoding (embedding) maps. We outline the relevant theory relying on differential geometry, and describe the corresponding gradient-descent optimization algorithm. The method is applied to standard data sets and we highlight its applicability, advantages, and shortcomings. In addition, we demonstrate that the same computational technology can be used to build coordinate invariance to local group actions when defined only on a (reduced) submanifold of the embedding space.

Conformal Disentanglement: A Neural Framework for Perspective Synthesis and Differentiation

For multiple scientific endeavors it is common to measure a phenomenon of interest in more than one ways. We make observations of objects from several different perspectives in space, at different points in time; we may also measure different properties of a mixture using different types of instruments. After collecting this heterogeneous information, it is necessary to be able to synthesize a complete picture of what is `common' across its sources: the subject we ultimately want to study. However, isolated (`clean') observations of a system are not always possible: observations often contain information about other systems in its environment, or about the measuring instruments themselves. In that sense, each observation may contain information that `does not matter' to the original object of study; this `uncommon' information between sensors observing the same object may still be important, and decoupling it from the main signal(s) useful. We introduce a neural network autoencoder framework capable of both tasks: it is structured to identify `common' variables, and, making use of orthogonality constraints to define geometric independence, to also identify disentangled `uncommon' information originating from the heterogeneous sensors. We demonstrate applications in several computational examples.