Why Accuracy Isn't Enough: A New Quality Model for Real-World ML Components

And that brings us to our quality models.

So let's talk about quality models.

So there are traditional software product quality models, like ISO 25010 is a standard that is widely used by industry.

More recently, there is ISO 25059, which is a quality model specifically for AI systems.

So Rachel, let's start with you.

What was the catalyst for developing yet another quality model, this time specifically for machine learning components?

So quite a few of these quality models start to mix system attributes with ML component attributes, which can make testing difficult as system attributes cannot be tested at the component level.

So by developing a quality model specifically for ML components, developers, acquirers, and stakeholders can use these QAs as a common vocabulary for generating requirements for the exact ML components and as a focus for testing efforts.

So additionally, in developing the new quality model, even though it is just one more, we were able to build on our data science expertise and make the new quality model both specific to ML components and relevant across the backgrounds that you typically see on a development team.

So to give you an idea, although this would be better explained visually, the quality model contains 30 quality attributes that describe system behavior that may be important to an ML component.

And so in your individual system, you might not use all 30, but you could.

At the highest level, the quality model is divided into seven categories.

So we have behavior analysis, confidence, consistency, continued operation, maintenance and evolution, responsible AI, and security and privacy.

then each one of these categories has two or more quality attributes under it, like a tree.

And so, for example, the confidence category contains explainability, functional correctness, and interpretability.

Yeah, so I think what you're saying, Rachel, is that one way to look at this quality model would be like, I don't want to say a checklist, but almost like if you're interacting with stakeholders, if you're trying to figure out, okay, what is important for my machine learning component, you could say, okay, do I care about robustness? Do I care about fairness? So is that a good way to think about it?

Yes, and you can also start to dive into exactly what those quality attributes mean for the exact system.

So you're not getting a mismatch on just one person says that robustness is important, another one says it's important, but they don't mean the same

thing.

Right, right.

Actually, that's a good point.

So getting consistency in the naming and the terminology that people

use.