A Bit About Supervised Learning
It is immediately obvious who is teaching here?
When I first heard the term “Supervised Learning”, I automatically thought it meant human feedback was a part of each and every result that a model would generate. I believe that the ‘supervised’ part of this terminology means that humans have already categorized the data that a model receives. So the “quality” of the output from the model isn’t automatically known in the process of Supervised Learning. More on that later.
Categorized Data vs. Uncategorized Data
This difference sounds so trivial, but a model built to receive already classified datasets will likely be very different from another that is meant to recognize opportunities to assign categories to data that has not already been categorized.
This could honestly be the work of two separate models. One has the job of finding unique correlations or unique opportunities for classification in a dataset, and another model does something with the categorized data.
Applications of Supervised Learning
From what I can glean, supervised learning is critical for applications where a precise output is needed. Identifying artifacts in images for example. If there was a Quality Analysis tool that received pictures of items from a production floor, its sole job in the world may be to give a quality score to that widget, and potentially identify defects.
It’s not up for debate that it must be right as close to 100% of the time as possible.
Applications of Unsupervised Learning
The applications for unsupervised learning are far more open-ended and less strict. Correlation seeking is the first thing that comes to my mind. Let’s use the previous example again.
Let’s say I had a batch of parts that were returned, and they all passed our ‘visual’ inspection from earlier. An unsupervised model could take all those images and see if it could find a common correlation between them. There is a chance it could identify a new defect that the original QA model wasn’t trained on. If this is the case, it would give a new ‘category’ of defect and that would be a part of the training data for the QA Model, and together they would help to create a more robust product.
Conclusion
Supervised and Unsupervised learning have very different capabilities but can be used together in concert if needed. Unsupervised learning is much more open-ended and is not expected to give predictable outputs. Supervised learning uses pre-classified data to create very predictable outputs and is likely the correct choice for “highly-repeatable” use cases.
