This project follows along with many modifications to the 4 year old tutorial of (details in document) about Deep Neural Networks and a short demonstration on his data not used here. We revisit the data set made from the PCA in Random Forest project earlier (see that project for link to data) and see how Keras can manipulate and solve the class imbalances of the data to make predictions on our 4 class target. There was some packaging and dependency issues and changes dealt with in a document not published but do know you should have Rtools installed, and latest keras and tensorflow to install those packages that are built for python modules but transitioned with an R package called reticulate that has some dependency issues I found but didn't publish but had to use nested for loops for the 4 classes. In the end, DNN does better on very large data and not 86X80 more like 860,000X80 as it was built for facial recognition and fingerprint matching, etc. The results are better than PCA using the error or noise to predict classes, but not better than random forest in the caret or randomForest package of R on this type of data.