03 Jun 2026
Shuo Presents His work at CVPR
On 3rd June 2026, Shuo is presenting virtually at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), held in Denver, Colorado. The event gathers leading academic researchers, industry pioneers, and engineers to discuss foundational methodologies and transformative artificial intelligence applications. With its rigorous standards and focus on cutting-edge computer vision architectures, CVPR serves as a premier international venue for presenting research that actively shapes the future of machine learning.
During the conference, Shuo presents his work, titled "Why Not Hyperparameter-Friendly Optimisation? A Monotonic Adaptive Norm Rescaling Approach For Long-Tailed Recognition". In this work, we propose Self-Adaptive Monotonic Normalisation (SAMN), a novel approach designed to tackle the significant challenge of long-tailed recognition. Specifically, it introduces a hyperparameter-friendly optimisation strategy that improves upon the popular two-stage decoupling paradigm. By utilising the Pool Adjacent Violators Algorithm (PAVA), the SAMN framework successfully enforces monotonicity on per-class weight norms without the need for traditional parameter regularisation.
SAMN demonstrates remarkable robustness by elegantly bypassing hyperparameter sensitivity, enabling stable training even in highly imbalanced datasets. Rigorously validated on benchmark datasets, this universal and plug-and-play strategy seamlessly integrates with other methods to significantly boost performance, offering a highly scalable new tool that consistently achieves state-of-the-art results.
