LeNet-5 Benchmarking: MNIST vs. CIFAR-10

A historical evaluation of the LeNet-5 CNN architecture (LeCun et al., 1998) trained using Stochastic Gradient Descent (SGD) on two foundational computer vision benchmarks.

🚀 Quick Results

Dataset	Type	Input Dim	Test Accuracy
MNIST	Handwritten Digits	$32 \times 32 \times 1$	96.65%
CIFAR-10	Real-world Objects	$32 \times 32 \times 3$	45.48%

💡 Key Findings

• MNIST (96.65%): LeNet-5 easily extracts spatial features from simple, single-channel geometric patterns.
• CIFAR-10 (45.48%): The shallow depth and low feature map capacity ($6/16$ channels) limit the model's ability to generalize across high intra-class color variations and complex backgrounds.

🛠️ How to Run

1. Clone the repository.
2. Install requirements: pip install tensorflow matplotlib scikit-learn datasets.
3. Run the main notebook: LeNet-5.ipynb.