New method prevents over-contraction in AI models for long-tailed recognition

Researchers have developed a new method for long-tailed recognition (LTR) that addresses the issue of tail classes being over-contracted in existing neural collapse (NC) approaches. Their technique, controlled partial neural collapse, uses Class-Adaptive Sparse Activation (CASA) and an ETF-constrained adaptive margin (ETF-AM) loss. This approach allows head classes to collapse while preserving within-class diversity for tail classes, leading to improved accuracy on datasets like CIFAR100-LT and ImageNet-LT. AI

IMPACT This research could lead to more robust AI models capable of better handling imbalanced datasets, improving performance in real-world applications with diverse data distributions.

RANK_REASON The cluster contains a research paper detailing a new method for improving AI model performance on long-tailed recognition tasks. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Casa
• CIFAR100-LT
• Class-Adaptive Sparse Activation
• Equiangular tight frames from group divisible designs
• ETF-AM
• ETF-constrained adaptive margin
• Full Collapse Is Suboptimal: Controlled Partial Neural Collapse for Long-Tailed Recognition
• ImageNet-LT
• iNaturalist-2018
• Places-LT

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →