Some things I'd like to see at WWDC 2021

So, WWDC starts on Monday, June 7, and I can unashamedly say it’s one of my most anticipated developer events of the year. I enjoy all Apple keynotes (despite sometimes being just a little too self-congratulatory), and the WWDC one is my favorite. It’s best when they focus on features in each of their new versions of iOS, macOS, iPadOS, etc., and at its worst when they use their platform to attack their competitors. So, as well as just focussing on their products, here’s a few things I would love to see at WWDC this year.

Create ML Model Interpretability

Create ML is a great tool, but for some reason, Apple does not give you any indication of the model architecture it creates. Additionally, the various workarounds that used to be available have now been obfuscated. That’s a big step backward in an era of model interpretability. Let’s be honest here, nobody creating a model with a tool like Create ML is developing a novel architecture – they’re making a model using existing images, and the tool creates the model, so what is there to protect? Worse – what if there are fairness issues with your model? Not having any control over its architecture is just crazy. So, please, Apple Developers, take it in the other direction and let folks see the architecture of the model output by the tool.

And let’s be honest here – in the case of Image Classification, it’s probably just a new classification head on the bottom of a model like MobileNet or EfficientDet, with the new number of classes. So why obfuscate? Or, best of all, maybe give us developers the opportunity to choose?

Likelihood: Low

Know your data - like tools in Create ML

On the theme of AI Fairness and ethics, it would be fantastic if a tool like Know Your Data got integrated into Create ML. It’s an immensely useful tool that lets you inspect your data against things that could cause possible biases. For example, please take a look at my Horses of Humans dataset. I created this explicitly to avoid bias by having men and women of multiple sizes, skin tones, etc. Without realizing it, in many cases, the lighting I used made humans of some skin tones have unrecognizable faces, and I would never have known this without KYD!

KYD Issue

So, if people create models quickly and easily with Create ML, let’s not lead them down the path of having biased models unintentionally! Please integrate KYD, or something like it!

Likelihood: Tiny

Export to TF Lite

TensorFlow Lite is the premier framework for running ML on all flavors of mobile. Android, iOS, JavaScript, Linux-based systems, embedded systems, etc. Create ML is a terrific tool. Wouldn’t it be nice for it also to export in .tflite format, so we could use it as a tool to create models for all of the above? I know it’s not the typical approach for Apple, but it sure would be nice!

Likelihood: Near absolute Zero

AirTags API

I bought a pack of 4 Air Tags and am waiting for my holders to attach them to things I don’t want to lose. They are undeniably cool, not least because they use the network of iOS devices globally to act as finders. There’s the possibility of applications beyond ‘Find My’ being able to use them. I have a game I want to write to use them instead of beacons, which didn’t live up to the promise. I’m hoping Apple will extend the Nearby API for Airtags at WWDC. Bonus points if their Nearby Interaction demo, which in my book goes down as one of the best developer demos ever, is updated for Air Tags.

NI Demo Image

Likelihood: Quite possible

Access to heart rate sensor in realtime

Maybe this is already available, but I can’t find it. With the watch having a heart rate sensor, it would be nice to have an API that triggers an alert when the heart rate goes above a certain amount. You might even be able to gamify this. Honest story: I was once playing Catan with my wife, and she was offering a trade that would have won her the game. It seemed a good trade to me. But, her apple watch was visible, and I saw her heart rate was higher than usual, indicating (to me, at least) that she was probably lying! I didn’t take the trade. It didn’t matter; she won on the next turn! But it did get me thinking!

Likelihood: Probably Not