OSDTW: Optimal Shared Depth and Task Weighting for Long-Tailed Recognition
Researchers have introduced OSDTW, a novel framework designed to tackle the long-tailed recognition problem in machine learning. This approach decomposes the recognition task into distinct head and tail components, utilizing a shared encoder with task-specific decoders. OSDTW provides a principled method for optimizing representation sharing and supervision weighting, offering a computable proxy for hyper-parameter selection based on a bias-variance decomposition of generalization error. Experiments on standard benchmarks show OSDTW outperforming existing methods. AI
IMPACT Introduces a principled framework for improving long-tailed recognition, potentially enhancing model performance in real-world scenarios with imbalanced datasets.