Predicting who you will call in the next five minutes – what Gurdeep meant

At the 2014 Lync Conference, Gurdeep Singh Pall, Corporate Vice President for Lync & Skype Engineering said this:

We can actually predict who you will be calling in the next five minutes.

What he said has been rolling around my head for a few weeks now. I wanted to try and understand more about what he meant.

Whilst I wasn’t able to find any specific research paper or announcement which correlated with what Gurdeep said, I think I know what he was talking about. If so, his words throw some interesting light on the future direction of some of the projects I’ve found.

The first thing I noticed as soon as I started to look around was how much amazing work Microsoft Research is actually doing, much of it unsung. Did you know, for example:

• Microsoft Research employs over 1,100 scientists and engineers
• Since being created in 1991, it now has 11 labs worldwide and has published over 6,000 peer-reviewed publications
• Information and Learning from MSR influences virtually every product Microsoft ships

Here’s a concrete example: the Windows Phone 7 Keyboard. If you’ve found that typing on the Windows Phone produced less than the normal amount of errors, or that the word predictions seemed to be getting it right rather than wrong more often, that’s because MSR has had a part in it. The principles of machine-learning improve the raw data being generated by your fingers. By combining statistical models of language patterns and touch points, the keyboard dynamically changes the virtual size of the likely next letter, so that it has a larger target area—the area where tapping the keypad results in a particular letter, symbol, or number. You don’t see this happening, the target areas are invisible – but it explains why your clumsy typing always seems to work out.

Microsoft Research is doing hundreds of other, equally fascinating things which I highly encourage you to go and have a look at – it’ll get you excited about things to come. From Elevators that predict which floor you want to go to, predicting Cholera outbreak, predicting your mood based on how you’re using your smart phone, or even predicting your future location.

However, there’s two projects in particular which I want to highlight, because I think they will impact the world of Lync the most, and will have the greatest potential.

Project Falcon

Project Falcon attempts to remove the delay inherant with opening mobile applications by predicting what application you’re going to open next, and pre-loading it in the background. That way, when you do open it it’ll open instantly as it’s already performed all its pre-loading steps.

If you can look past the (presumably royalty-free) lift music, there’s a demonstration video of Falcon in action:

Or if that’s too much for you, The Verge did a piece on this.

Falcon seems to be part of larger project, ConDOS. The idea behind ConDOS is that that operating system should do a much better job of aggregating and presenting situational data. So, for instance, rather than the OS exposing that the GPS has moved backwards 0.5M, the accelerometer has shown downward then stationary activity, ConDOS would signal to applications “the user is sitting”.

By taking this legwork out of data analysis things become much easier for application developers. Things like fast launching become more possible, and it enables loads of other scenarios.

The reason I bring this up? I think that knowing what application you’re about to open is impressive, but it’s only Part One. Part Two would be knowing why you’re about to open it, what you’re about to do. If you know this then you can pre-load information inside the application, making things even faster. Maybe this is what Gurdeep meant by knowing who you’re about to call. Maybe Project Falcon has moved onto the next stage already?

During the Keynote, Gurdeep said:

This is about context, it’s about intelligence. This what I was I was doing for the last two years. I was building an AI platform which allows applications to use and deposit context which can then be used by other applications. This is really powerful.

From the CondOS Research Paper:

We propose the design of a new mobile phone operating system … that embraces context generation. CondOS provides both applications and internal OS services the ability to query or subscribe to CDUs [Context Data Units] while protecting raw sensor data privacy…In complement, CondOS provides a way for apps to extendthe CDU vocabulary by defining their own CDUs for inclusion in the kernel.

This might not be what he meant, but even if it isn’t – this is still going to enable some pretty amazing scenarios.

Bayesphone

The second piece of research I want to highlight is the academic sounding Precomputation of Context-Sensitive Policies for Inquiry and Action in Mobile Devices.

Sounds dull, but actually it’s one of the coolest things I’ve seen in a long time.

The project had two aims:

1. trying to predict whether users will attend meetings on their calender, and
2. if they’re in a meeting, what the cost of an incoming call during that meeting would be, and whether it should be allowed to ring or not

The project uses Bayesian probability models to address these problems. (if you use a decent spam filter then there’s a good chance it uses Bayesian probability to weed out the dodgy emails from the good ones).  These are real situations and problems and I think it’s brilliant that Microsoft is paying people to come up with solutions for them!

Oh, and because I know you’re dying to know what the magic formula for calculating the cost of interrupting someone during a meeting:

You’re going to have to read the research paper to get the meanings of all those variables!

Summary

I’m going to finish with something else Gurdeep said during the Keynote:

“This is going to transform, not just Communications, this is going to transform everything in the Enterprise.”

You’d better be ready.