Neural Net - MNIST

It is a university project and an exercise for neural networks.

Data is obtained from this website in CSV format. The activation function is sigmoid, and the weight distribution has been derived from the following relationships: $\pm1/\sqrt{incomin Links}$ A normal distribution where the mean is zero.

Model Accuracy and Optimization

The model has achieved an accuracy of 97.53% with a learning rate of 0.07, over 5 epochs, and 600 nodes. It exhibits the lowest accuracy of approximately 95% on digits 7 and 2, while the highest accuracy is around 99% on digits 1 and 0.

For new data, the model requires optimization as it has been trained. It currently struggles with noisy images.

Note on Optimization

To maintain the integrity of the model's performance on new datasets, it is crucial to optimize it following the established training protocols.

Requirement

• matplotlib==3.8.4
• imageio==2.34.1
• numpy==1.26.4
• scipy==1.13.0

Artikel - in persian language

Note

You can test the network with new data, but because the network is trained with clean data and the number recognition algorithm is not implemented, it cannot recognize noisy data, and gives wrong output.

site on Streamlit cload