Jearmy said at the end of lesson 10 to continue investigating what happens inside the model when training at high lr with the activation plots and histogram and to try and increase the accuracy on MNIST higher than 98% in 8 epochs using our insights.
I think this is a great idea cause instead of just understanding what happens inside we are able to derive practical insights to better training. I would like to know what people observed and came up with to increase the accuracy.