The Role of Pseudo-Labels in Self-Training Linear Classifiers on High-Dimensional Gaussian Mixture Data
The episode discusses a paper analyzing self-training classifiers using pseudo-labels on high-dimensional data. Hosts explain that while self-training is effective, its mechanism is complex. They detail how optimal strategies change based on iteration count and propose two heuristics—pseudo-label annealing and bias-fixing—to improve performance in label imbalanced settings.