Reproduce “Blind Source Separation of Single-Channel Mixtures via Multi-Encoder Autoencoders” on OpenResearch

About this paper

Matthew B. Webster, Joonnyong Lee

A multi-encoder autoencoder is introduced by researchers from Mellowing Factory Co. Ltd. for blind source separation of single-channel, non-linear, and underdetermined mixtures, employing a fully self-supervised approach with novel regularization. The method effectively disentangles sources in a synthetic dataset and extracts respiration from raw ECG and PPG signals, achieving respiratory rate Mean Absolute Errors (MAEs) of 1.50-1.73 breaths/min, which is superior to established heuristic techniques.

Key insights

• Deep neural networks, specifically multi-encoder architectures, can exhibit feature subspace specialization, which can be harnessed for implicit source separation without explicit supervision.
• Novel regularization techniques, like sparse mixing loss and zero reconstruction loss, are effective in guiding the self-supervised separation of complex, non-linear sources within a single-channel mixture.
• The method successfully tackles highly challenging BSS problems by learning separation directly from the data distribution, moving beyond strong prior assumptions about the mixing process.

Summary by alphaXiv.