ResLT

ResLT addresses long-tailed visual recognition by operating in parameter space with a residual fusion architecture to improve classification of underrepresented (medium and tail) classes.

Key Features:

• Parameter Space Perspective: Focuses on adapting the parameter space rather than re-sampling or re-weighting input or loss spaces to preserve model capacity for low-frequency classes.
• Residual Fusion Mechanism: Implements a main branch plus two residual branches where the main branch recognizes all classes, the first residual branch enhances medium and tail class recognition, and the second residual branch specifically targets tail class improvement.
• Additive Shortcuts: Aggregates outputs from the main and residual branches via additive shortcuts to combine branch contributions into final class predictions.

Scientific Applications:

• Bioinformatics: Improves classification accuracy for biological datasets with long-tailed class distributions by enhancing detection of underrepresented classes.
• Image recognition: Enhances performance on long-tailed image classification benchmarks by allocating dedicated residual capacity to rare classes.
• Ecological studies: Improves species recognition in ecological datasets containing many rare taxa, such as those with long-tailed occurrence distributions.

Methodology:

Evaluated on long-tailed versions of CIFAR-10 and CIFAR-100, Places, ImageNet, and iNaturalist 2018; experimental results on these benchmark datasets demonstrate the effectiveness of the parameter-space residual fusion approach for long-tailed recognition.