Overview

This project compares Multi-Layer Perceptrons (MLPs) and Convolutional Neural Networks (CNNs) on image classification tasks using MNIST and CIFAR-10.

The project investigates:

Model performance across increasing dataset complexity Parameter efficiency Inductive bias in CNNs Training dynamics and computational efficiency Technologies Used Python PyTorch NumPy Matplotlib scikit-learn Datasets MNIST CIFAR-10 Experiments Part A — Complexity Ladder Fixed MLP architecture Fixed CNN architecture Comparison across datasets Training curves and accuracy analysis Part B — Parameter Budget Challenge Models constrained to ~500K parameters Hyperparameter tuning Efficiency analysis Key Concepts Inductive bias Spatial locality Translation invariance Parameter efficiency Results

CNNs significantly outperformed MLPs on CIFAR-10 due to their ability to exploit spatial structure in image data.

Author

Kamva Poswa University of Cape Town