Machine Learning (ML) is an extremely powerful tool that has been distinctly out of reach for the average person since its conception. For my fellow laymen, ML is the method of training a computer model to classify something. That’s it. Whether it’s supervised or unsupervised, all these programs are doing is recognizing and memorizing patterns from examples. While a high-level explanation is easy to understand, its implementation is far from it. However, what if we took that simple understanding and, acknowledging that what I’m about to say will make armies of developers and engineers wince, picked out specific topics we want to classify and then just build models for those?
Those who have taken advantage of ML within their businesses know that it’s not nearly that simple. Building and optimizing a model takes time. A lot of time. Gathering training and testing data is time-consuming, build times can exceed days depending on the size of your model, and updating the model once it has been built means feeding in more data. Heaven forbid you realize that the original question you were trying to answer was slightly off, requiring you to start from scratch.
This pain however isn’t true for unstructured language. Language is much, much harder. So why even entertain the thought “what if we just pick some stuff to classify”, especially in the context of unstructured text? Because that’s exactly what Gracie does.
Gracie has allowed me, a somewhat technical person with a degree in Philosophy, to build machine learning models that interpret and classify unstructured text with a high degree of accuracy, without any programming. The only skill required is the ability to read. Models that myself and other non technical users have built range from document type classifications (W2, obituary, RFP, etc.) to more abstract concepts such as opaque language, patient frailty, and urgency.
The secret to Gracie’s elasticity and accuracy lies in what we call human augmented machine learning. As the human in this equation, I know when something is opaque and what a W2 looks like, allowing me to tune Gracie based on my domain expertise. As the machine, Gracie’s models perform according to the training material, and the results are curated to refine her understanding. The marriage of ML and human guidance allows users to drastically cut down on the amount of training text required, while patented technologies make the build times almost instantaneous and classification proximities intuitive.The end result is a balanced system of input and output where both Gracie and the user agree on what constitutes ‘yes’ and ‘no’.