I am trying to create a prediction model which would return a numerical score upon receiving a set of representative numbers. The purpose was to replicate human quality assessment procedure upon facing visual and textual data. In my prototype, I've made a python script to extract representative data from the raw data (video and text) and trained with the human assessment score. However, due to the small assessment data size (~60 pts), it was impossible to account for every possible variation (i.e. having low accuracy on predicting what was not trained). Therefore, my solution was to train using a set of statistical distribution curve/function from the existing dataset, then the trained machine accuracy would be polished by having a second round of active learning from a small number of real humans. I've developed the first prototype on Keras, but am willing to create a more serious version this round. What are your suggestions in terms of implementation? Also, What are your thoughts on using a statistical distribution curve to train the NN and polishing with active learning?