Neural network: Using genetic algorithms to train and deploy neural networks: Experiment - Compare FPP with TensorFlow


(0 comments)

We do experiments to compare TensorFlow - A popular machine learning program that trains networks based on loss functions, and FPP - Network training based on genetic algorithms.

The goal is not to compare the two software, but to propose the method of network training by genetic algorithms instead of relying entirely on loss function.

Both solve a problem. That is "Text classification with movie reviews". Both use a sample set with the same network configuration.

Below is the configuration table



And below is the table of training parameters and results


The results show that FPP has a higher test accuracy, but more importantly it has much lower training accuracy than TensorFlow. This proves that FPP uses Genetic Algorithms overcoming the overfitting better than other methods, or it is more generalizable.


Here are some pictures


The problem "Text classification with movie reviews" with TensorFlow. 40 epochs. The later the test accuracy decreases. No good number of epochs to be determined.




The problem of "Text classification with movie reviews" with FPP. The training process automatically stops when the Genetic Algorithms converge. The number of epochs is determined automatically, in this case 17.

You can download the file fppdata.txt, then run the program like this:

./fpp -l --units-min=4 --units-max=4 -j4 -t10000 -x

 

Currently unrated

Comments

There are currently no comments

New Comment

required

required (not published)

optional

required


What is 9 + 7?

required