In the chapter about “Understanding Text” a FFN is implemeneted that outputs an accuracy of 0.7676411 on the test dataset in page 213. Then the book moves on to implement an RNN that outputs an accuracy of 0.7510081 on the test dataset in page 219. I was a bit puzzled since the expectation and description is that the RNN would perform better but the results proof otherwise.
Is this intentional? Or perhaps the examples in the book are missing some updates.
There is actually an issue with training RNNs in Axon right now Training them is unstable and so results in much worse results than you would normally see. I’m working on fixing the issue, you can track the progress here: Performance diversion between RNN in Keras and Axon · Issue #530 · elixir-nx/axon · GitHub
Thanks for the prompt response @seanmor5 . Makes sense, I will be following on the github issue, seems like a great way to get introduced more in-depth in the inner workings of the libraries.
I will take the chance to congratulate you for the book. I am new to machine learning and I am enjoying it a lot. Thank you and cheers!