implementation of a neural net from scratch to classify digits(0-9) using the MNIST dataset. I also made a GUI to make it more interactive.

digit recognizer - Hello World.ipynb loads the MNIST dataset, builds a NN from scratch and saves weights-biases to various files so this NN can be reconstructed in the app.py

app.py reconstructs the NN, displays an interactive GUI.

layers : 784x128x128x128x10 hidden layers uses "relu" as an activation function output layer uses softmax activation function. loss gets calculated with categorical cross entropy loss.

weights and biases initializes randomly

weights = 0.01 * np.random.rand(self.layers[i], self.layers[i + 1])

biases = 0.01 * np.random.rand(self.layers[i + 1])

learning rate = 0.01, epochs = 10

accuracy is around 97%

demo

user's drawing

resulted image after converting canvas to 28x28

Thus, the neural network itself is working with 97% accuracy, however my application is not perfect :(