Noisy Labels Leave Geometric Signature in Neural Networks, Study Finds

Researchers have developed a method using persistent homology to analyze how noisy labels affect neural network representations. Their experiments on image and text classification tasks, including CIFAR-10 and CIFAR-100 datasets with ResNet-18, revealed that corrupted labels leave a distinct geometric signature in neural activations. This signature, measurable through loop structure analysis, becomes apparent even when the network's predictions align with the noisy labels only a small fraction of the time, offering insights beyond standard training accuracy metrics. AI

IMPACT This research provides a novel method to understand how noisy data impacts model learning, potentially leading to more robust training techniques.

RANK_REASON The cluster contains a research paper published on arXiv detailing a new methodology for analyzing neural network representations. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• CIFAR-10
• CIFAR-100
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv
• ResNet-18
• ScienceCast

• paper
• model release

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