Skip to main content
Back to Projects
AI/MLMar 2026 – Present

CIFAR-10 Image Classification

An end-to-end machine learning pipeline comparing custom CNN architectures trained from scratch against transfer learning with pretrained models (MobileNetV2, ResNet-18, EfficientNet-B0, ViT) on the CIFAR-10 benchmark. Includes advanced data augmentation, Grad-CAM interpretability, INT8 quantization for deployment, and an interactive Streamlit demo.

Tech Stack

PythonPyTorchscikit-learnStreamlitJupyterGrad-CAM

Key Highlights

  • Designed and implemented an end-to-end image classification pipeline to evaluate five deep learning architectures on the CIFAR-10 dataset
  • Compared training-from-scratch and transfer learning approaches under controlled conditions, with MobileNetV2 reaching 85.53% test accuracy and outperforming a custom CNN by nearly 30 percentage points
  • Applied modern training techniques, including data augmentation, cosine annealing learning rate scheduling, and progressive unfreezing, to improve convergence and generalisation
  • Extended the project with Grad-CAM interpretability, INT8 quantisation, command-line inference tools, and a Streamlit interface to support analysis and deployment