All projects

Hybrid Deep Learning for Noisy Label Classification

An applied machine learning system for image classification under noisy real-world labels, using contrastive pretraining and robust learning strategies.

PyTorchComputer VisionContrastive LearningRobust Learning

Overview

This project explores building reliable image classification systems when training data contains noisy or imperfect labels — a common real-world challenge in applied machine learning.

What I Built

  • ResNet-based image classification pipeline
  • Contrastive pretraining workflow
  • Robust training strategies for noisy labels
  • Reproducible training and evaluation system

Key Engineering Decisions

  • Used contrastive pretraining to stabilize representation learning
  • Focused on robustness and generalization instead of benchmark chasing
  • Designed experiments for reproducibility and structure
  • Prioritized system reliability over raw accuracy metrics