Machine Learning Checklist

Do you have a good application for machine learning?

Not all applications or business processes can be improved with machine learning, and it isn't a panacea for any issue. There is a lot of hype around artificial intelligence (AI); even the name is misleading. AI doesn't have any intelligence at all, at least in the sense that a human has intelligence. It simply a mathematical model that was trained on a bunch of data that with known inputs and outputs, and then when presented with a new input, generates a corresponding output. It isn't able to say if the output makes sense, nor can it come up with anything truly novel. It can only combine bits of what it was trained on, and interpolate between those bits of data. For example, you could show it the design for every piston engine made for an airplane, and it would be able to come up with new piston engine designs (say a bigger engine, or more cylinders, etc). But it could never conceive of a turbine engine. AI works that way. Despite some amazing abilities, thinking outside of the box is not one of them.

Disclaimer aside, how can you tell if you have an application that can be improved with machine learning? Currently, the most robust part of machine learning concerns images - both still and video. If your business process involves images, then it is quite likely machine learning can help. There are two basic operations that machine learning does on images - it can detect objects and give you a bounding box of their location in the image, and it can determine what pixels in the image belong to what object. The following questions are a quick list that can help determine how well suited machine learning is to your business process:

• Do you have a way of collecting images in your business process (e.g. smartphones, industrial cameras, etc).
• Can a human identify the features, items, changes that you want to detect?
• Do you have a way to collect a large quantity of data with both the inputs and outputs identified?
• Can your process support a step where a human reviews the results?
• What is the risk or cost of a mis-identification?

To work through the list, let's pick a simple application. Suppose that you supply temporary staffing to hospitals and other medical facilities, and that your workers need to get signed timesheets when they complete their work. In order to streamline the process, you might allow the workers to use their smartphones to photograph the timesheets and send them to you for processing. Your staff will reject any timesheet that doesn't have two signatures - one for the employee and one for the client. Now let's go through that checklist:

• Clearly there is a way to collect the images, as essentially all professional workers have a smartphone.
• A human can easily see if there are two signatures on a form.
• If the business has been running for a while, then there is a large body of signed timesheets. This can serve as the basis for training data. Also, in this case, because the application is very common (signatures are a big part of a lot of business processes), there are already trained models freely available.
• This process already has a 100% human review, so there is no issue with having someone review the results. That review is very important, because as good as machine learning is, there will always be at least some small amount of errors. No model is perfect.
• In this case, a mis-identification could cause a client to not pay for the work that was done. But because of the human review step (i.e. once machine learning says that there are two signatures, a human confirms that), there is no additional risk.

So for this example, using machine learning can reduce the amount of time to process timesheets, and improve the rate of catching incorrectly signed ones.

I would love to talk to you about your business process, and see if machine learning might be useful. Please contact us if we can be of some help.