A comparative study of calibration methods for imbalanced class incremental learning

Deep learning approaches are successful in a wide range of AI problems and particular for visual recognition tasks. However, there still open among which is the capacity to handle streams information management class imbalance datasets. Existing research these two separately while they co-occur real world applications. Here, we study problem incrementally from imbalanced We focus on algorithms have constant deep model complexity use bounded memory store exemplars old classes across incremental states. Since bounded, learned with fewer images than new an due added initial dataset imbalance. A score prediction bias favor appears evaluate comprehensive set calibration methods reduce it. Evaluation carried three datasets, using configurations sizes. Results show that most beneficial effect useful lower sizes, interesting practice. As secondary contribution, remove usual distillation component loss function algorithms. simpler vanilla fine tuning stronger backbone