Fixed Non-negative Orthogonal Classifier: Inducing Zero-mean Neural Collapse with Feature Dimension Separation

FNO classifier makes the LPM achieve the global optimality even in inducing the max-margin decision while satisfying the properties of the zero-mean neural collapse and invokes feature dimension separation, which is useful in continual learning and imbalanced learning

Recommended citation: Kim, H. & Kim, K. (2024), Fixed Non-negative Orthogonal Classifier: Inducing Zero-mean Neural Collapse with Feature Dimension Separation. The Twelfth International Conference on Learning Representations
Download Paper | Download Slides