The abstract for this arXiv paper introduces a problem in representation learning for time series, using human motion data as a motivating example. It states that such recordings often encode structure across multiple temporal scales, and when a latent state determines the sequence, this structure needs to survive in the representation space. The sentence ends mid-thought in the posted abstract, leaving the proposed mechanism and results unspecified.
Judging from the title, the intended contribution is a method based on local temporal contrasts that induces a latent geometry from human motion data. The idea appears to be that contrasting nearby time points can help learn representations where the multi-scale structure is preserved as geometric properties. However, because the abstract is truncated, no further details about the method, experiments, or conclusions are available in this source.
This digest is based solely on the single arXiv listing, so there are no differing sources to compare. Readers should consult the full paper for the missing content.