Neural Network

Implementation of Tariq Rashid "Make Your Own Neural Network" using Python.

How does the neural network work in a nutshell

Setup

Docker

You can use the Dockerfile to start a container running Ubuntu with all required packages installed.

Mnist Data

To train the network you should download full train data https://www.pjreddie.com/media/files/mnist_train.csv and testing data https://www.pjreddie.com/media/files/mnist_test.csv

curl https://www.pjreddie.com/media/files/mnist_train.csv -o data/mnist_train.csv

curl https://www.pjreddie.com/media/files/mnist_test.csv -o data/mnist_test.csv

Add both files to the /data folder or execute: python download_mnist.py

Benchmark

To run the benchmark execute python Benchmark.py

You can configure the benchmark by changing the following constants:

• number of epochs
• hidden_nodes_list
• learning_rate_list

Best performance: 0.9783

• epoch: 10
• hidden_nodes: 400
• learning_rate: 0.05

Tasks

• Destroy single or a specific amount of nodes/weights to see how much it affects performance.

• Visualize the weights between the layers.

• Send output signal reverse to get an image back from the input nodes.

• Make the number of hidden layers and their nodes configurable.

• Try different activation functions and see how they change the performance.

• Rotate the train images by +/- 10° to improve performance.

Notes

• How does the number of hidden layers affect the performance?