AI at Work

14 Jul 2017 . category: Career . Comments

A serendipitous thing (but also a thing that’s been a long time coming, for me personally, and for a Microsoft product in the Satya Era) has happened recently. My org at Microsoft has decided to give a small group of engineers, product managers, and designers some space and time to investigate what Artificial Intelligence could mean for our note-taking products. I am fortunate enough to have joined this effort.

It’s a new frontier for this particular team, and I’m extremely excited to have this new focus at work so nicely coincide with the personal learning that I will be detailing on this blog. Whenever I want to blog about work-related data science challenges and learnings, I will aim to generalize the details enough to be applicable widely.

We are hoping to develop a sort of brain trust for the org that will eventually make itself obsolete. For the next few months, we will be taking a look at the AI landscape, external and internal to Microsoft. We will form opinions on what techniques have potential and what may not, particularly for notes. We will distill and translate the theory and the buzzwords of the field into practical knowledge. We will wade into the compliance, user privacy, and data protection discussions happening throughout the company. And then, we will disseminate what we have learned so that the entire org can bake intelligence into every feature.

Personally, I have three main work goals for this fiscal quarter in AI. I hope to implement an automatic task (ToDo) detection system, a smarter OneNote picker a way to find related notes1, and an automatic system for learning and improving a OneNote feature based on user feedback (perhaps with reinforcement learning?)2. This is ambitious for a single quarter, but I am engaged and ready to give it my best.

What a time to be alive.

Footnotes

  1. Update 8/23/2017: My teammate has picked up the endeavor of figuring out how to build a section classifier that would benefit the OneNote picker scenario, which is great! I’ll be able to chime in and contribute to that work. Related notes is the next on my list. ↩

  2. Update 8/31/2017: I’ve learned some things since this and now realize that reinforcement learning isn’t what I wanted here. I actually want a system that supports active learning for the interactive querying of users and online learning for the incremental updating after each user query (or batch of user queries). Reinforcement learning is still a cool concept though! ↩

Me

Nadja does not particularly enjoy writing about herself.