neuralnet-from-scratch

2026 · NumPy, PyTorch · MNIST, EMNIST · Source

Draw a digit from 0 to 9.

What the
network sees

Draw in the box to see the network’s guess.

Neural network guess.

How it works

A 784 → 16 → 16 → 10 network with ReLU hidden layers and a softmax output, trained with plain SGD (learning rate 0.1, batches of 100, 10 epochs) on the 60,000 MNIST training digits. Your drawing is cropped, scaled and centred.

10,000 MNIST test images, mean ± standard deviation over 10 runs.
Implementation	Test accuracy
NumPy, from scratch	95.00 ± 0.27%
PyTorch, same network	95.07 ± 0.31%

How it works

A wider 784 → 128 → 128 → 47 network trained on EMNIST Balanced: 112,800 handwritten digits and letters. Lowercase letters that look like their capitals (c, o, s, x and others) share a class with them, which leaves 47 classes. Your drawing is cropped, scaled and centred.

18,800 EMNIST test images, mean ± standard deviation over 10 runs. In each run every optimiser starts from the same weights and sees the same shuffle order.
Optimiser	From scratch	torch.optim
SGD, learning rate 0.1	82.91 ± 0.30%	82.99 ± 0.29%
SGD with momentum 0.9, learning rate 0.01	82.89 ± 0.34%	82.90 ± 0.30%
AdamW, learning rate 0.001	83.33 ± 0.24%	83.34 ± 0.19%

Each from-scratch optimiser matches its PyTorch counterpart within run-to-run variation. Some characters are ambiguous even to people, such as 0 and O, or 1, I and l, so expect the probability to be split between them.