A simple CNN (Convolutional Neural Network) built from scratch using Java (plus some UI).

Dataset: MNIST - hand-written numbers (Training set: 60000, Test set: 10000). I used the CSV version of the dataset: https://pjreddie.com/projects/mnist-in-csv/

Implemented batch training because the training speed and training accuracy was too low on the whole dataset and I didn't know how to use GPU on Java.

Accuracy after 10 epochs was 0.85 with batch size of 6000 and a simple architecture: CONV, POOL, FULLYCONNECTED (reLU).

TO BUILD: just run the main.java.

References:

• Rea is Online: https://www.youtube.com/watch?v=3MMonOWGe0M&list=PLpcNcOt2pg8k_YsrMjSwVdy3GX-rc_ZgN
• Victorzhou: https://victorzhou.com/blog/intro-to-cnns-part-1/