Plasticity-Stability Preserving Multi-Task Learning for Remote Sensing Image Retrieval

Deep learning-based multi-task learning (MTL) methods have recently attracted attention for content-based image retrieval (CBIR) applications in remote sensing (RS). For a given set of tasks (e.g., scene classification, semantic segmentation, and reconstruction), existing MTL employ joint optimization algorithm on the direct aggregation task-specific loss functions. Such an approach may provide limited CBIR performance when: 1) compete or even distract each other; 2) one dominates whole procedure; 3) characterization task is underperformed compared to single-task learning. This mainly due lack of: plasticity condition (which associated with sensitivity new information) stability protection from radical disruptions by procedure. To avoid this issue, as first time, we propose novel plasticity-stability preserving (PLASTA-MTL) ensure conditions procedure independently number type tasks. achieved defining two The function (PPL) that aims enforce global representation space be sensitive information learned task. minimizing difference gradient magnitudes embedding spaces. second (SPL) protect radically disrupted angular distances between gradients over space. effectively proposed functions, also introduce sequential algorithm. Experimental results show effectiveness state-of-the-art context CBIR.