lenet5-digit-demo

A small LeNet-5-style convolutional network for handwritten digit classification (0-9), used in an interactive browser visualization that shows a hand-drawn digit flowing through each layer of a real, trained CNN.

Model Details

• Architecture: LeNet-5 -- 6@5x5 conv, pool, 16@5x5 sparsely-connected conv, pool, FC(120), FC(100), FC(10). Tanh-family activations throughout.
• Input: 28x28 grayscale image, drawn by the user, normalized to roughly [-0.1, 1.175].
• Output: 10-way digit classification.

Training Data

Trained on JustinCignal/mnist-lenet5-demo, a repackaged MNIST training split.

Training Procedure

Trained for several epochs with Adam (lr=1e-3), MSE loss against scaled one-hot targets matching the network's tanh output range.

Intended Use

Demo / education only -- a lightweight interactive visualization of a CNN's internal layer activations while classifying hand-drawn digits. Not intended for production digit recognition.

Evaluation

~99% accuracy on held-out MNIST test images.