Spaces:

LaelaZ
/

nn-from-scratch

Sleeping

App Files Files Community

nn-from-scratch / README.md

LaelaZ

Add short_description (card/search one-liner)

16298b9 verified 5 days ago

preview code

raw

history blame contribute delete

870 Bytes

A newer version of the Gradio SDK is available: 6.17.3

Upgrade

metadata

title: Neural Network From Scratch (NumPy)
emoji: 🧠
colorFrom: blue
colorTo: indigo
sdk: gradio
app_file: app.py
pinned: false
license: mit
short_description: 'Neural net by hand in NumPy: 97.7% on MNIST, no framework.'

Neural Network From Scratch (NumPy)

A multilayer perceptron written entirely by hand in NumPy: every forward and backward pass, the softmax cross-entropy, and the Adam optimizer. No PyTorch, no TensorFlow.

It reaches ~97.7% accuracy on MNIST, and its hand-written backprop is verified against finite-difference gradients in the test suite (so the chain rule is provably wired up correctly, not just "it seems to train"). Draw a digit, or load a real MNIST test image.

Architecture: 784 → 256 → 128 → 10, ReLU activations, He initialisation, Adam.

Source & full docs: https://github.com/LaelaZorana/nn-from-scratch