On the Role of Inductive Bias in Time-Series Pretraining: A Case Study in Learning Generalizable Representations for Clinical Time Series

In three linesPathoFM, an encoder-centric transformer pretrained on clinical time series (pathological gait analysis for spinal cord injury), combines three objectives: Local Completion, Temporal Continuity, and Unsupervised In-Context Dynamics. The study shows that dynamics-centric objectives produce the most balanced transferable representations across classification and regression tasks.

Summary generated by Claude — human-verified