Opportunistic Teacher Forcing: Training RNNs on Wearable Signals with Inherent Missing Data

Sutashu Tomonaga, Andres Hernandez-Matamoros, Haruo Mizutani, Kenji Doya

Abstract: Wearable devices enable longitudinal monitoring of human physiology but suffer from inherent missing data, hindering analysis. Traditional recurrent neural network (RNN) training assumes complete datasets, while imputation introduces biases that corrupt the signal’s underlying dynamics. We propose Opportunistic Teacher Forcing (OTF), an extension of Scheduled Sampling that inherently handles missing data without imputation or architectural modifications. OTF switches probabilistically between teacher-forced and autonomous modes, critically computing loss only on available timesteps. We evaluate OTF using a Multi-Timescale RNN (MT-RNN) on surrogate data (Lorenz 63 system with up to 70% bursty missing) and real-world wearable data (ECG-like signals with ∼30% missing). Results demonstrate a robust reconstruction of the system’s geometry (via time-delay embeddings) and spectra (via Hellinger distance), validating OTF as a promising method for dynamical system reconstruction from incomple

External IDs:dblp:conf/biostec/TomonagaHMD26