New 'Noisier' NCE method improves density-ratio estimation for AI models

Researchers have developed a modified Noise Contrastive Estimation (NCE) technique called "Noisier" NCE, which addresses limitations in estimating density ratios for complex datasets. By artificially increasing the noise magnitude, this method aligns NCE gradients more closely with Maximum Likelihood Estimation (MLE), enabling faster convergence and improved performance. The approach has shown success in image modeling, anomaly detection, and offline optimization, achieving state-of-the-art results on datasets like ImageNet64x64 with significantly reduced training iterations. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Improves density-ratio estimation for complex datasets, potentially enhancing performance in image modeling and anomaly detection.

RANK_REASON Academic paper introducing a novel modification to an existing machine learning technique.

Read on arXiv cs.LG →

• Noise Contrastive Estimation
• Maximum Likelihood Estimation
• CIFAR-10
• ImageNet64x64

• paper
• other