Welcome to Your New OME

Apple CreateML interface showing model results

"Service guaranteed in one hour, or your money back!" When I'm used to "service" being completed within seconds, this was a surprise. But a new OME eventually took shape!

That is, an Orthogonal Model of Emotions. Even when I was testing IBM Watson it didn't take an hour to train and build the OME from less than 10k sources. I built it just to learn how Watson worked, but I learned more about Machine Learning and more sophisticated modeling.

However, it takes something as simple as what Apple provides for me to really start to tinker under the hood. When building an OME for demonstration or testing in an iOS environment, I could build and train a model using 10k worth of training data and then spin through multiple analyses of 10k RottenTomatoes reviews within mere minutes. When I increased the complexity of the classifier, I lost some of the accuracy. But it was still super fast! And mostly accurate. That should work for emotions, right?

But I've pushed the envelope with the CreateML app instead of just using Swift. It took about an hour, but I think the results of using extracting features using dynamic embedding in a transfer learning algorithm will be a game changer. Tokens, representing words, are best fit within approximate models within a model. Although the resulting OME TextClassifier is still simple, it has increased in size from 80KB to 1.5MB. I can see a measurable difference in the resulting analysis but I'll have to work with it some more to see what has changed.

Besides taking forever, that is! It's slow as my own thinking, it seems.