This repo contains the model I created for the MNIST-Kaggle competition, achieving 99.6 % accuracy in my best submission on the site.
The model and repo were made as an exercise after I participated in an introductory course in Deep Learning (in the winter semester 2019) to see if I could use all the techniques covered myself when creating a model from scratch. The model created uses the most basic building blocks in Deep Learning for Computer Vision like:
- CNN:s
- Batch Norm-layers
- Dropout-layers
- Data Augmentation
The network itself is rather big for the task and far from fully optimized, but achieves good performance on such an easy task as the MNIST dataset.