search

attractivechaos / kann

A lightweight C library for artificial neural networks
Other
678 stars 117 forks source link

Why accuracy does not increase anymore at specific epoch? #42

Closed co7ding closed 3 years ago

co7ding commented 3 years ago

Given MNIST CNN example, validation cost does not increase from approximately epoch 11. So, running more epochs is useless since validation cost will be only increasing or decreasing in near minimum cost value which is the validation cost of epoch 11. Could you explain why this happens and how to solve it? (I also tested using a variety of structures of CNN, but there was no big difference.)

attractivechaos commented 3 years ago

You have reached the capability of the model. You can't do much unless you change the model.

co7ding commented 3 years ago

I tried many models by changing the structure, but all structures that I tried could not reach an accuracy of over 90%. Did you get a high accuracy than 90% in CNN example?

attractivechaos commented 3 years ago

Just run the command lines in "examples". You can trivially get 97% validation and test accuracy even with MLP. Even one round of mnist-cnn gives you 2.6% validation error rate.

co7ding commented 3 years ago

Oh, I misunderstood the mnist-eval.pl code. Sorry for a stupid question. Thank you :)