This is a complete neural network implementation in plain C++ to solve the MNIST handwritten digit classification problem.
Run make to build the executable, then make train to begin training/testing. Use make run to run the prediction service using a Crow server.
Try the live demo
Neural network with 256 hidden neurons and 0.05 learning rate...
Epoch 1 | Training 100% | 95.29% Accuracy
Epoch 2 | Training 100% | 96.2% Accuracy
Epoch 3 | Training 100% | 96.53% Accuracy
Epoch 4 | Training 100% | 97.06% Accuracy
Epoch 5 | Training 100% | 97.21% Accuracy
Epoch 6 | Training 100% | 97.39% Accuracy
Epoch 7 | Training 100% | 97.62% Accuracy
Epoch 8 | Training 100% | 97.54% Accuracy
Epoch 9 | Training 100% | 97.65% Accuracy
Epoch 10 | Training 100% | 97.69% Accuracy