search

amalF / Kervolution

Kervolution implementation using TF2.0
20 stars 2 forks source link

unexpected results! #10

Open catproof opened 3 years ago

catproof commented 3 years ago

this is more of an issue with the original paper than it is with your repository, but I did some experimentation that showed that the results in the original paper claiming the KNN performs better than the CNN on the MNIST dataset are misleading. I did a learning rate search for both networks and showed that the CNN performs better given that they both use optimal learning rates! check out my repository here: https://github.com/NickPerezCarletonUniversity/Kervolution I have an in-depth analysis showing the results

catproof commented 3 years ago

i also tried this on a similar dataset and got the same results! :O