A new encoder for biomechanical movement sequences
Researchers implemented a Recurrent Temporal Restricted Boltzmann Machine to map movement data to fixed-length embeddings.
Researchers implemented the Recurrent Temporal Restricted Boltzmann Machine (RTRBM) as an unsupervised encoder. This model maps windows of movement data to fixed-length embeddings.
Biomechanical movement recordings are continuous, high-dimensional time series. The RTRBM is a recurrent energy-based model. It was adapted to continuous-valued signals and stacked into a deep architecture.
The model was integrated into the SIT-FUSE segmentation pipeline. It was tested on simulated musculoskeletal reaching and two real gait datasets. Testing included clustering, reconstruction, and walking-speed classification of unseen subjects.
Clustering the embeddings recovers movement structure tied to progress toward a target without labels. On gait data, the embeddings outperform PCA and autoencoder baselines under a nonlinear probe.
Evidence for a benefit of recurrence is mixed. Removing the recurrent connection does not yield a measurable loss on the reaching task. Simple per-channel summary statistics classify walking speed at least as well as the embeddings.
Key facts
- The RTRBM maps movement data windows to fixed-length embeddings.
- Clustering embeddings recovers movement structure tied to target progress.
- The model was tested on simulated reaching and two real gait datasets.
- Simple summary statistics classify walking speed as well as the embeddings.