Deep Learning

Image Classification using AlexNet on CIFAR-10

A deep-learning CNN that classifies 60,000 images across 10 categories.

202584.6% Test accuracy

OVERVIEW

The Project

This project implements a convolutional neural network based on the AlexNet architecture to classify images from the CIFAR-10 dataset of 60,000 colour images across 10 categories. The aim was a high-performance CNN that balances generalization and training efficiency.

Objectives

Implement an AlexNet-style model using PyTorch.

Achieve over 80% accuracy on the CIFAR-10 dataset.

Apply data augmentation and regularization techniques.

Tools & Technologies

PythonPyTorchMatplotlib & scikit-learnCUDA GPUCIFAR-10 dataset

FIGURES

Charts & Graphs

Confusion matrix — CIFAR-10 classification with AlexNet

METHODOLOGY

The Approach

Input CIFAR-10 images → data augmentation (random crop, flip, colour jitter).

AlexNet CNN: 3 convolutional blocks (ReLU, BatchNorm, MaxPooling) → fully-connected 1024 → 512 → 10.

Training: CrossEntropyLoss, Adam (LR 0.001, weight decay 1e-4), StepLR scheduler, 50 epochs, batch size 128.

Evaluation via accuracy and a confusion matrix.

OUTCOME

Results & Learnings

Final training accuracy ~90%, test accuracy 84.62%.

Overfitting reduced via augmentation and dropout.

Confusion matrix showed balanced class-wise accuracy.

Key Learnings

Data augmentation improves generalization significantly.
Classic architectures like AlexNet still perform well.
Visualization tools are vital for model diagnostics.

Related Case Studies