CNN on MNIST Data

---

Link to colab : https://colab.research.google.com/drive/1wFDEw5Y09pdLBZxq5hgP7W3FPTuUdlBz

Target:

Analyse the data

Create a basic skeleton of model using expand and squeeze model with GAP layer followed by FC layer(using 1x1) so that we get high accuracy in less than 10k parameters.

Results:

Parameters: 9,876

Best Train Accuracy: 98.56%

Best Test Accuracy: 98.58%

Analysis:

Data Analysis:

mean and standard deviation are 0.1307 and 0.3081 respectively.

some images were slightly rotated, some shifted, some quite distinguishable while others were marked with light marks. These transformatins can be used for image augmentation.

With 10k parameters, training up to 15 epochs led to test accuracy of around 98%.

We see that the starting accuracy is very low for such a small dataset like MNIST. This starting train/test accuracy has to be improved to reach higher accuracy within 15 epoch.