Biomechanical movement recordings arrive as continuous, high-dimensional time series. Analyzing them without manual labeling, the authors argue, requires representations that capture how movements evolve over time rather than static snapshots.
The paper, posted on arXiv, introduces Recurrent Temporal Restricted Boltzmann Machines as unsupervised encoders for these sequences. The model is designed to learn such temporal representations directly from unlabeled movement data.
Because this is the only source and the abstract is brief, the write-up is limited to the stated motivation and proposed approach; no experimental results or comparisons are reported in the available text.