## about ##

This is my attempt to create a digit recognition
neural network from scratch (without machine learning
frameworks, only numpy + math).

It consists of one input layer (for the 28 * 28 grayscale
pixel values), 2 hidden layers (each with 10 nodes) and
one output layer (which corresponds which digit is the most likely).

It uses relu activation layers for the hidden layers and
a softmax activation layer for the output layer.

It uses the MSE as the loss function. The error in this case is the
model output subtracted by a one hot vector of the expected output.

It uses standard backpropagation with a constant learning rate and
stochastic learning to update the weight matrix.

The weights are initialized using a normal distribution, as described
in the research paper of Le Cun et al.

I could get this model accurate up to 96%.


## dependencies ##

- numpy
- matplotlib


## running ##

`python3 main.py`

This will initialize the model and start training. Periodically a matplotlib
window will be opened showing a sample of test images and the model prediction
to visualize the model's accuracy.

----

This project is licensed under the MIT license

(c) 2024 Lucas Birkert