Innovation-July-August-2023

F E A T U R E

IS IT A LOW- OR HIGH-RISK APPLICATION?

There is some middle ground to be found in this area in “locked

a known function to the data and spits out an output, then it’s very clear how we would validate that function,” he says. There is also uncertainty about how it will behave if given new data that it wasn’t trained on. That initial data set is vital, he says, as well as ensuring it has all the real-world inputs that the AI might see. All the questions Henrey raises are active areas of research, and there has yet to be a consensus on how to validate AI. One technique is to validate the output rather than trying to validate the tool itself. “I think that’s a good way of building confidence in a tool. Do the same work that the tool is doing in something representative and make sure that the outputs are correct,” he says. “But there’s a big caveat here. There is the potential for that underlying function to change over time in ways that we wouldn’t expect.”

algorithms,” he says, where the model is trained on data, then frozen so it will behave the same way moving forward. However, using an algorithm that changes, learns, and adapts over time, makes it much more difficult to assure consistency. “If that algorithm is processing some sort of safety critical sensor data, then we need to have confidence that it’s going to process the safety critical sensory data correctly today, tomorrow, and a year in the future,” he says. On the other hand, if underlying functions are unknown—which can be the case with a lot of tools that have been trained on an unclear dataset—it can be a lot more difficult to have good confidence. There’s not really a clear way of obtaining that confidence, he adds. “If we have a traditional algorithm where we feed it data and it’s applying

An important question to ask is: what is the intended use of the tool? That helps guide the risk process. “The effort that goes into validation needs to match with the amount of risk that the tool is associated,” says Henrey. High-risk usually involves AI integrated into a product, perhaps sensing real-world parameters, he says. Henrey says widespread use of ChatGPT put AI into the mainstream. “It’s made that transition from being a software engineering tool to a tool that is accessible to pretty much everybody in the public now,” he says. “I think we’d probably find that we’ve already incorporated low-risk tools into our day-to-day life, for example with project planning and scheduling.” While Engineers and Geoscientists BC has guidelines that are applicable to AI (particularly paragraph 3.3.3. in the BC Guide to the Standard for Documented Checks of Engineering and Geoscientist

GEOLOGICAL ENGINEERING LIBRARY

FREE BOOKS

265 BOOK LIBRARY

TOPICS

Anchoring Blasting Dams

Mathematics Mining Rock Falls Rock mechanics Rock Slopes Tunnels Risk management

Earthquakes Foundations Geology Ground water

Contact Duncan Wyllie www.wnrockeng.com > library for catalogue of books

1 4

J U L Y / A U G U S T 2 0 2 3

I N N O V A T I O N