TL;DR: Using Discriminative Training of Kalman Filters (2005) to your tune filter’s process noise.

I recently was asked about how you tune the noise covariance matrices for (Extended) Kalman Filters. It was a pointed question, where I felt obliged to have an answer, and while I had a good answer for the measurement noise $R$, I could only remember iteratively hand tuning $Q$ based on some educated guesses and perceived performance. I assumed there was a better answer, and I was embarrassed I didn’t have it, so I set out to find it.

Right: Rudolf Kalman, inventor of the Kalman Filter, receives the National Medal of Science from President Obama. Left: Margaret Hamilton, one of the leading minds who developed the Apollo on-board flight software, sits in the Apollo module. The navigation system interface is above her head, below the circular attitude indicator, running Kalman’s filter.

Three topics, here’s the TL;DR!

1. I created a cheatsheet for SLAM terms, resources, and important papers. (SLAM stands for Simultaneous Localization and Mapping, see the glossary!) It’s a work in progress.
2. Working through ORB SLAM, I realize it is misnamed. While ORB features are important part of the technique, they could be swapped out for another feature-detector-descriptor. The greater innovation seems to be in the thoughtful weaving of many algorithms through four threaded processes making map updates.
3. Alarming survey results from consulting biostatisticians were released (Annals of Internal Medicine) showing that researchers frequently make inappropriate requests of biostatisticians to alter results. One stat: 24% of respondents had been asked to “remove or alter some data records to better support the research hypothesis”. Mark Twain comes to mind …

I go in depth on each below the fold …

I miss blogging, and I thought it would keep me focused to publish a short post about what I’m learning, or what projects I’m working on, each week.

Google cars have driven 2 million miles in self-driving mode (if you believe in projections.) And Google is making a huge push to drive more autonomous miles.

This is short tutorial on setting up Caffe to run on Amazon Web Services (AWS). If you’re hacker who lives on the command line, this is not for you. If instead, you are a recovering pure mathematician who is uncomfortable with the messy innards of your laptop, and even more uncomfortable with the cloud, this might be for you. You can find other tutorials with similar content, for example in Stanford’s CS231 course on Neural Networks. This might be more of a meta-tutorial, in that, for some cases, I plan to point you to resources and say “do this”.

¡Bienvenido!

I’m excited to start blogging again, now as a Data Scientist, on this blog: AlexThompson.ai . I’ve been transitioning to data science over the last 14 months, and it is time to get a bit more public about it.