One of the great pleasures of my position at Catapult Systems is to be able to share my excitement about artificial intelligence technologies with a broad range of people. I get to do this through our Approachable AI workshops which we deliver around the United States. Wednesday, October 16 saw us in our corporate hometown of Austin, Texas. We had a good variety of industries represented, including Oil and Gas, Semiconductors, Software, as well as a good contingent from the City of Austin!
I wanted to take the chance to note down a few of the key things I hope people took away from the day:
1. The Math
You don’t need to understand all of the mathematics underlying machine learning algorithms to be able to leverage this technology. In fact you can start to build and deploy simple models without even knowing how to code using Microsoft ML Studio. That said, to be effective you do need to understand the steps in the data science workflow. You also should have a feel for how tuning the hyperparameters of a machine learning model may influence its performance.
2. Models Are Naïve
As of today, our machine learning models can only learn patterns from the training data we put in front of them. If we haven’t taught a computer vision model to recognize a star, it will not be able to do so. And if we give a model biased data, the model will learn that bias. It is down to us as practitioners to ensure our data is complete enough for our use case, and that we go to great lengths to ensure bias is not present – directly or indirectly (eg through a zip code).
3. Automate or Augment?
Machine learning can be used to automate a process or to augment a human’s decision making, or both. But it is never a requirement to cede complete control over to a machine, and the appropriateness of doing so depends on the use case. Maybe we would do it for an email marketing campaign, but almost certainly not for criminal justice purposes.
4. Set Good Expectations
Setting good expectations at the start of a project is the single biggest thing you can do to maximize your chances of a successful project. For example, no model for a problem worth solving with AI will actually have 100% accuracy.
5. Beat the Baseline!
All your first model has to do is beat the current system. Is the machine learning model more accurate than your current statistical model, or can it perform a task faster than a human for the same error rate? Iterative improvement can realize business value faster, and will eventually yield a valuable model.
I hope all the participants found the workshop valuable and interesting. And if you want help identifying and prioritizing the problems at your organization that machine learning might solve, please reach out and let’s chat about our Innovation Workshop.