Design Under Uncertainty

Graph of a bivariate normal distribution

With the recent explosion in the capability and reach of machine learning, we must become literate in both its strengths and its limitations. Today’s ML-based systems aren’t really Artificial Intelligence per se — they’re just special-purpose pattern-matching machines. If we feed them biased data to inform their inference engines, they’ll produce biased results. Even with the best-intentioned, most balanced data models, mis-predictions or mis-identifications still happen — simply due to normal probabilistic variation. But how does the user understand these errors? This class of smart system typically can’t explain itself when it makes a dumb mistake, because its algorithms don’t lend themselves well to self-diagnosis. We need to make progress on this problem because explainability helps build trust.

Contemporary, ML-backed translation software works on a sequence-to-sequence basis; it translates whole phrases or sentences rather than individual words. As design lead for the Siri Translations feature, I had to deal with the constraints that this back-end imposed. Namely, out-of-context reverse-translation of individual words was taken off the table.

Siri Translations
Siri Translations

From mid-2017 to mid-2019 I was the interaction design lead on Siri’s Advanced Development Group, where I focused (among other things) on machine-learned gestural interactions such as Raise to Speak for Apple Watch, which launched in watchOS 5. This feature allows you to omit the “Hey Siri” trigger phrase when you raise the watch all the way to your lips.

Raise to Speak for Apple Watch
Raise to Speak for Apple Watch