How the fuck do people get 98.5% on this
I've tried CNNs with # of parameters ranging from 100k to 6m with batchnorm, dropout and data augmentation via adam and it is literally impossible to get above 96% accuracy
what am I missing
>>60720708
Probably with terabytes of input and slightly more time on their hands then you have.
Cherry pick your test images.
Its all about the face of the animal, between cat and dogs they're very distinct shapes.
All you need to do is apply some simple shape matching.