Considerations when validating and auditing algorithms

The future is now. Industry 4.0 probably means you have algorithms in your process. For example, if you aren’t using algorithims to analyze deviations, you probably soon will.

 And with those algorithms come a whole host of questions on how to validate and how to ensure they work properly over time. The FDA has indicated that ““we want to get an understanding of your general idea for model maintenance.” FDA also wants to know the “trigger” for updating the model, the criteria for recalibration, and the level of validation of the model.”

Kate Crawford at Microsoft speaks about “data fundamentalism” – the notion that massive datasets are repositories that yield reliable and objective truths, if only we can extract them using machine learning tools. It shouldn’t take much to realize the reasons why this trap can produce some very bad decision making. Our algorithm’s have biases, just as human beings have biases. They are dependent on the data models used to build and refine them.

Based on reported FDA thinking, and given where European regulators are in other areas, it is very clear we need to be able to explain and justify our algorithmic decisions. Machine learning in here now and will only grow more important.

Ask an Interesting Question

The first step is to be very clear on why there is a need for this system and what problem it is trying to solve. Having alignment across all the stakeholders is key to guarantee that the entire team is here with the same purpose. Here we start building a framework

Get the Data

The solution will only be as good as what it learns from. Following the common saying “garbage in, garbage out”, the problem is not with the machine learning tool itself, it lies with how it’s been trained and what data it is learning from.

Explore the Data

Look at the raw data. Look at data summary. Visualize the data. Do it all again a different way. Notice things. Do it again. Probably get more data.  Design experiments with the data.

Model the Data

The only true way to validate a model is to observe, iterate and audit. If we take a traditional csv model to machine learning, we are in for a lot of hurt. We need to take the framework we built and validate to it. Ensure there are emchanisms to observe to this framework and audit to performance over time.