Category Archives: Machine learning

Quick thoughts: on Apple’s subtle machine learning improvements

John Gruber had a post on Tuesday about the improvements to Siri over time. He was mostly talking about the core functions of Siri itself, but I’ve noticed some other subtler improvements over the past couple of years that I’ve been meaning to write up for ages. Technically, these aren’t part of Siri per se, but they sit in the same category as Cortana or Google Now’s equivalent functionality, so I see them as almost an extension of Siri. What’s striking about these to me is that as far as I can tell Apple hasn’t explicitly talked about them and since they don’t trigger explicit notifications many users might not even notice them.

I grabbed a screenshot of an example of this back in February 2014, and I think it’s the best way to explain the sort of thing I’m talking about:

2014-02-08 07.08.45It’s the first paragraph there that I’m focusing on, and I need to provide some context to enable you to understand what happened here. Early last year I played for a few weeks in a church basketball tournament. I had games regularly each Saturday at a similar time, and at the same location. In my calendar I put simply “Basketball” and never entered the address, since I knew already how to get there. What Apple’s machine learning engine did here was (as far as I can guess 1):

  • Note that I had an item called “Basketball” in my calendar for that morning
  • Make a connection with past appointments on Saturday mornings also called “Basketball”
  • Look up past location behavior in its location database to connect a particular location with past instances of “Basketball” in my calendar
  • Look up this address and calculate driving time between my current location and this destination
  • Present it to me at a relevant time in the Today screen.

Again, Apple has talked up some functionality around using calendar locations explicitly entered in your calendar to provide these sorts of alerts, but I’m not sure it’s ever talked about the deeper machine learning stuff in evidence here. I’ve never seen exactly this sort of extrapolation from past behavior again since this occasion, but I have received other notifications on this screen that it’s time to leave for appointments where I’ve explicitly entered a location in my calendar, based on heavy traffic (it happened to me this past week at CES, for example).

I wonder how long it will be before Apple starts surfacing this information more proactively with notifications rather than merely displaying it reactively when users pop open this screen. And I wonder if Apple will begin talking about this functionality more in future – I note that Microsoft is currently running adds which state that Siri can’t tell you if you need to leave for a meeting based on traffic. That’s technically true – Siri is entirely reactive today – but iOS does tell you if you know where to look. That seems like something Apple could easily fix, but it would turn its machine learning capabilities from a background feature that’s so subtle many users likely miss it into a headline feature. And with that might come concerns about  the data Apple is gathering and how it’s using it. It may have decided that, for now at least, it would rather keep quiet about all this, but I wonder how long that will last.

Notes:

  1. Another possibility is that Apple merely built a pattern of my past behavior on Saturday mornings without explicitly connecting them to the calendar item. I don’t see that it makes a big difference either way.