FISHER AI framework boosts rare fish species ID by 13.4%

FISHER tackles the fundamental challenge of negative transfer in multi-task learning (MTL) for fine-grained aquatic species recognition. In traditional MTL, gradient conflicts between low-level dense tasks (e.g., segmentation) and high-level classification objectives degrade feature representations, especially for ultra-rare species with few training examples. The framework enforces a unidirectional information flow from segmentation to trait prediction to species classification, explicitly decoupling gradients across task boundaries. This prevents high-level objectives from corrupting low-level morphological representations while preserving shared supervision benefits.

FISHER also introduces a prototype-based segmentation head with orthogonality regularization to encourage disentangled anatomical representations, and uses homoscedastic uncertainty weighting to dynamically balance task contributions during training. The analysis shows that robust trait representations serve as critical bridges for transferring knowledge to ultra-rare species. On the Fish-Vista benchmark, FISHER achieves 97.7% mean average precision (mAP) for unseen trait identification and a 13.4% improvement in ultra-rare species classification accuracy over strong baselines, demonstrating the effectiveness of gradient-decoupled hierarchical learning for long-tailed biodiversity recognition. This work has significant implications for automated biodiversity monitoring, fisheries management, and conservation efforts where accurate identification of rare and morphologically similar species is essential.