Signal reconstruction by means of Embedding, Clustering and AutoEncoder Ensembles

We study the denoising and reconstruction of corrupted signals by means of AutoEncoder ensembles. In order to guarantee experts' diversity in the ensemble, we apply, prior to learning, a dimensional reduction pass (to map the examples into a suitable Euclidean space) and a partitional clustering pass: each cluster is then used to train a distinct AutoEncoder. We study the approach with an audio file benchmark: the original signals are artificially corrupted by Doppler effect and reverb. The results support the comparative effectiveness of the approach, w.r.t. the approach based on a single AutoEncoder. The processing pipeline using Local Linear Embedding, k means, then k Convolutional Denoising AutoEncoders reduces the reconstruction error by 35% w.r.t. the baseline approach.