A new theoretical framework for neural collapse (NC) has been developed, addressing challenges posed by imbalanced datasets and open-set scenarios. The research reformulates NC using class-wise averaged losses instead of global ones, proving that standard NC is achieved at the minimum of these class-wise losses for both cross-entropy and binary cross-entropy. The study also extends this to open-set conditions, showing that classifier vectors for unknown classes overlap and become orthogonal to known classes, while standard NC persists for the known classes. Furthermore, the paper demonstrates that re-balanced Softmax- and Sigmoid-based contrastive losses also exhibit standard NC at their minima on imbalanced datasets. AI
IMPACT This research provides a theoretical foundation for understanding and improving classification models in challenging data scenarios.
RANK_REASON The cluster contains an academic paper detailing theoretical research on neural networks.
Read on Hugging Face Daily Papers →
- alphaXiv
- arXiv
- Binary Cross-Entropy
- CatalyzeX
- cross entropy
- DagsHub
- Gotit.pub
- Hugging Face
- IArxiv
- neural collapse
- ScienceCast
- sigmoid function
- Softmax
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
