Noise$^+$2Noise: Co-taught De-noising Autoencoders for Time-Series Data

Harry Rubin-Falcone, Joyce Lee, Jenna Wiens

Keywords: De-noising, Co-teaching, Noise recovery, Time-series, self-supervised, RNN

TL;DR: We combine Co-teaching and De-noising Autoencoders to recover clean signals from only noisy data in a time series setting.

Abstract: We consider the task of learning to recover clean signals given only access to noisy data. Recent work in computer vision has addressed this problem in the context of images using denoising autoencoders (DAEs). However, to date DAEs for learning from noisy data have not been explored in the context of time-series data. DAEs for denoising images often rely on assumptions unlikely to hold in the context of time series, \textit{e.g.}, multiple noisy samples of the same example. Here, we adapt DAEs to cleaning time-series data with noisy samples only. To recover the clean target signal when only given access to noisy target data, we leverage a noise-free auxiliary time-series signal that is related to the target signal. In addition to leveraging the relationship between the target signal and auxiliary signal, we iteratively filter and learn from clean samples using an approach based on co-teaching. Applied to the task of recovering carbohydrate values for blood glucose management, our approach reduces noise (MSE) in patient-reported carbohydrates from 72$g^2$ (95\% CI: 54,93) to 18$g^2$ (13,25), outperforming the best baseline (MSE = 33$g^2$ (27,43)). We demonstrate strong time-series denoising performance, extending the applicability of DAEs to a previously under-explored setting.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Machine Learning for Sciences (eg biology, physics, health sciences, social sciences, climate/sustainability )