Filter Guided Manifold Optimization in the Autoencoder Latent Space

Abstract

An autoencoder is a class of neural network that is trained to output an accurate reproduction of the input while learning key lower dimensional features, otherwise known as a manifold. A lower dimensional representation of the original input, referred to as the latent space, encodes the intrinsic data structure over the manifold. This paper proposes filter-guided manifold optimization in the latent space of a convolutional autoencoder to recover noisy motion data collected by a depth sensor. Autoencoder output is smoothed using four traditional filters and employed as target motion data in an objective function. The difference between the actual output and target is minimized through stochastic gradient descent over the latent space, using manifold optimization to produce the expected smooth output. The advantage of this filter-guided approach over traditional filtering is that the resultant motion data still adheres to the manifold in the latent space learned by the autoencorder from training on motion data.