Machine unlearning strategies compared for noisy label correction

A new study published on arXiv explores the effectiveness of various machine unlearning (MU) strategies for correcting noisy labels in deep neural networks. Researchers compared five MU methods—NegGrad, Fine-Tuning (FT), Random Labeling (RL), SalUn, and MUNBa—across different types of noise, including symmetric, asymmetric, instance-dependent, and open-set noise. The findings indicate that the optimal unlearning strategy depends on the specific noise structure, with simple fine-tuning performing well in most closed-set scenarios. Robust methods like RL and SalUn approached retraining accuracy at a significantly lower computational cost, especially under instance-dependent noise. AI

IMPACT Provides practical guidelines for selecting machine unlearning strategies to improve model performance when dealing with noisy data.

RANK_REASON Academic paper detailing a comparative study of machine unlearning methods for a specific task. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• CIFAR-10
• CIFAR-100
• Fine-Tuning
• Food-101N
• Machine Unlearning
• NegGrad
• Random Labeling
• SalUn

• paper
• other

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