Lack of support for contrastive learning across multiple patients, which is required by some contrastive learning pretraining tasks. NCL is the one I have in mind. These pretraining methods often use the supervised label or dynamic time warping within a batch, so maybe we can add a stage for batch mining.
Proposed Solution
Patient ID in the batch: Include patient_id in the window data struct to allow for tracking the patient and mining positive and negative pairs
Implement Multi-Patient Contrastive Learning
Problem
Lack of support for contrastive learning across multiple patients, which is required by some contrastive learning pretraining tasks. NCL is the one I have in mind. These pretraining methods often use the supervised label or dynamic time warping within a batch, so maybe we can add a stage for batch mining.
Proposed Solution