Project 2 in FYS-STK4155, fall of 2018. Made in collaboration with Betina, Ingrid, Polina and Mona. More details may be found in the project report.

In this project we study and attempt to make an comparison between our own NumPy-based CNN, Keras CNN, TensorFlow CNN and support vector machine model of MNIST database. We find that to make a computationally fast CNN is quite hard. However, prediction accuracy for NumPy-based CNN, Keras CNN,TensorFlow CNN and support vector machine models are 95.3%, 98.9%, 99.0% and 96.9%, respectively. Due to computational difficulties of NumPy-based CNN we had to resort to a smaller database of digits, 8 x 8 database, to obtain such a good accuracy. Another aspects that we could emphasize in case of NumPy-based, TensorFlow and Keras CNNs is that employed structure of CNN model and the training has an immense impact on the final prediction accuracy.