Neural Network MNIST Classifier

A C++ feedforward neural network trained on the MNIST handwritten digit dataset, with a raylib GUI for drawing digits and running live predictions. The trained saved weights and biases have an accuracy of 95.4%, though you can probably get a higher accuracy by playing around with the layer sizes and changing the condition on line of neural_net.cpp : if (epoch_accuracy>best_accuracy+0.003).

The network is implemented from scratch using Eigen.

Flow Chart

Video Explanation

Features

• Custom MLP with configurable layer sizes, ReLU hidden layers, and softmax output
• Cross-entropy training with mini-batch gradient descent (batch size 32)
• Early stopping on a held-out validation split with best-weight checkpointing
• Model persistence — save and load weights to data/best_model.bin
• Interactive GUI — draw digits with the mouse, preprocess like MNIST, and classify in real time
• Background training — model loads or trains on a worker thread while the window stays responsive

Architecture

Default network: 784 → 256 → 64 → 10

Input:28×28 grayscale image Hidden layers:ReLU function Output layer:softmax function Loss:Cross-entropy Optimizer:stochastic gradient descent

Training uses 48000 MNIST images and 12000 are left for validation. Test accuracy is reported on the official 10000 image test set.

Hand-drawn digit preprocessing

To match MNIST normalization image processing:

1. Capture the 280×280 drawing canvas
2. Find the bounding box of white ink pixels
3. Crop to that box
4. Scale to fit inside a 20×20 region (aspect ratio preserved)
5. Paste into a 28×28 black image, centered by center of mass at (13.5, 13.5)
6. Feed the resulting vector to the network

The right panel in the GUI shows the processed 28×28 input the network actually sees.

Build

cmake -B build
cmake --build build

The executable is build/main.

Run

./build/main

You can choose to either load a saved model from data/best_model.bin or train your own network by editing line 23 of src/main.cpp: NeuralNet net({784, 256, 64, 10}, 32, true, true). You can change the layer size list (first entry must be 784, last must be 10), batch size (32), and the two flags at the end. Set save_read to true to read weights from file on startup (skipped if the file is missing or the layer list does not match), or false to always reinitialize and train; set save_write to true to overwrite and save data/best_model.bin on quit when training improves validation accuracy, or false to leave the file unchanged.

Wait for "Training not done, please wait" to disappear before submitting a drawing.

Controls

Left mouse drag: Draw on the canvas (white on black) Enter: Preprocess drawing and run prediction C: Clear the canvas Q: Quit (saves model weights on exit)

Predictions appear below the processed-image preview on the right.

Model file format

data/best_model.bin format is as follows:

[float]  best validation accuracy
[int]    number of layer sizes
[int]    layer_sizes[]        
per layer:
    [double] weights
    [double] biases