Hybrid Deep Learning for Noisy Label Classification

Overview

This project implements and compares four label-noise robustness strategies — cross-entropy baseline, neighbor-consistency regularization, agreement-based hybrid reweighting, and contrastive pretraining — under a fixed ResNet-18 pipeline on CIFAR-10N human-annotated labels.

What I Built

• ResNet-18 classification pipeline with four noise-robustness strategies
• Structural and hybrid loss functions combining feature-space neighbor consistency with confidence-based sample reweighting
• Contrastive pretraining workflow to stabilize representation learning
• Diagnostic evaluation pipeline: learning curves, confusion matrices, UMAP/t-SNE embeddings

Key Engineering Decisions

• Achieved 87.53% peak validation accuracy on CIFAR-10N vs. 86.91% baseline and eliminated 2.4pp of late-epoch memorization decay
• Designed hybrid loss combining neighbor consistency and confidence-based reweighting to address both noise and representation geometry
• Used UMAP/t-SNE to interpret confusion patterns across visually-ambiguous classes
• Focused on reproducibility with a fixed pipeline to isolate the effect of each robustness strategy