- New Scientist did a short article on my PhD work.
- As of September, 2012, I've received my PhD, and work at Facebook. I'm more likely to update my blog than this site, so check there for updates.
- A journal version of our work on generating storylines from sensor data was published in Pervasive and Mobile Computing, available here
- My thesis, "Efficient computational methods for understanding human behaviour from mobile phone data", has been accepted, and is available here
- Our paper, "Generating storylines from sensor data", placed 2nd in the Open Track at the Nokia MDC Workshop at Pervasive 2012. Thanks to Nokia for hosting a fantastic workshop! Paper now available here.
- For a quick 3-minute overview of our work on gait and activity recognition, watch this video from the ECML-PKDD 2011 demo session.
- I presented our work at the AAAI Spring Symposium on Computational Physiology. The talk [SLIDES] covered some exciting recent results that have been submitted and are under review for publication.
- We are making available a large real-world data set for the Gait Recognition task. We hope to spur good empirical comparisons between competing algorithms in this area. If you use this data in published work, please include this citation.
- Code for Time-Delay Embedding feature extraction (Frank et al., 2010) now available under MIT license.
- Initial release of our Android Data Collection Platform. Open source (MIT License) framework that facilitates easy development of applications that either produce or consume data on the Android Platform.
- The Discovery Channel did a piece on some of our group's work. Watch it here (our bit starts at 8:24).
- Our paper, Activity and Gait Recognition with Time-Delay Embeddings, was accepted for publication in the AAAI 2010 conference.
My name is Jordan Frank, and I'm a software engineer at Facebook. The information on this page is related to work undertaken during my graduate studies in the Computer Science department at McGill University in Montreal, Quebec. I was supervised by Profs. Doina Precup and Shie Mannor. My research was supported, in part, by the NSERC Alexander Graham Bell Ph.D. Canadian Graduate Scholarship, for which I will always remain extremely grateful. Here is my relatively current CV. I am a member of the Reasoning and Learning Lab and the Center for Intelligent Machines. I'm interested in many topics, from compilers to cryptography, but my research area was in, what I like to call, ubiquitous machine learning. The big question that motivates my work and studies is:
What could we learn if we had access to all of the data collected by all of the electronic devices that every person on earth interacts with on a daily basis?
To answer such a question requires scaling machine learning algorithms up to deal with a barrage of incoming data from a complex, non-stationary world. I can think of no better place in the world for me to continue working on answering this question than at Facebook.
- Generating storylines from sensor data (Pervasive and Mobile Computing), with S. Mannor, and D. Precup. August 2013 (BibTex).
- Efficient computational methods for understanding human behaviour from mobile phone data. PhD Thesis. McGill University. November 2012.
- Generating storylines from sensor data, with S. Mannor, and D. Precup. Mobile Data Challenge 2012 (by Nokia) Workshop, in conjunction with the International Conference on Pervasive Computing, June 2012. (second prize in the Open Challenge track).
- Time Series Analysis Using Geometric Template Matching (IEEE Transactions on Pattern Matching and Machine Intelligence), with S. Mannor, J. Pineau, and D. Precup. March 2013 (BibTex).
- Activity Recognition With Mobile Phones (ECML 2011 Demo Session), with S. Mannor and D. Precup, Sept, 2011 (Video).
- Mobility Profile and Wheelchair Driving Skills of Powered Wheelchair Users: Sensor-Based Event Recognition Using a Support Vector Machine Classifier (EMBC 2011), with A. K. Moghaddam, J. Pineau, P. S. Archambault, F. Routhier, T. Audet, J. Polgar, F. Michaud, and P. Boissy. Aug., 2011.
- A Novel Similarity Measure for Time Series Data with Applications to Gait and Activity Recognition (UBICOMP 2010, adjunct proceedings), with S. Mannor and D. Precup, Sept. 2010.
- Activity and Gait Recognition with Time-Delay Embeddings (AAAI 2010), with S. Mannor and D. Precup, July 2010. BibTex
- Learning state space models from time series data (short abstract for MSRL 2009).
- Reinforcement learning in the presence of rare events (Masters Thesis). McGill University, August 2008.
- Reinforcement learning in the presence of rare events (ICML 2008), with S. Mannor and D. Precup (Video, BibTex).
- Recognizers: A study in learning how to model temporally extended behaviors (NIPS 2007 Workshop on Hierarchical Organization of Behavior: Computational, Psychological and Neural Perspectives) with D. Precup. (Slides from presentation are here).
- Time-series analysis using time-delay embeddings. Talk at 2011 AAAI Spring Symposium on computational physiology [SLIDES].
- Talk on Reinforcement Learning in the Presence of Rare Events at ICML 2008.
- I have presented versions of our mobile phone app during the demo sessions at ECML-PKDD 2011, UBICOMP 2010, and NIPS 2009.
- HumanSense Android Data Collection Platform. App for the Android platform that allows flexible logging and analysis of sensor data. Also includes demo of activity recognition and location discovery from wifi signals.
- Time-Delay Embedding Feature Extraction Tools. Some command-line tools for building classifiers based on time-delay embeddings of time-series data.
- Gait Recognition Data Set. Collected from sensors in a Nexus One mobile phone. 20 subjects walking outdoors for 15 minutes on two different days. If you intend to publish results on this data set, please cite: Jordan Frank, Shie Mannor, and Doina Precup. Data Sets: Mobile Phone Gait Recognition Data, 2010 [BibTex].
My PhD work is focused on a project that we are calling the The Digital Rashomon Project. From Wikipedia:
The Rashomon effect is the effect of the subjectivity of perception on recollection, by which observers of an event are able to produce substantially different but equally plausible accounts of it. We are currently investigating methods for modeling human behaviour using data collected from smartphones. More information can be found here. We have developed a system called WalkID that can identify the person carrying a mobile phone based on their gait, or style of their walk. The Discovery Channel did a piece on this technology, which can be seen online here (our segment starts at 8:24). See the Publications section above for more details.
We have recently developed and made available an opensource framework for developing data-driven applications on the Android platform. The framework consists of a set of classes that can be used for developing plugins that either consume or produce data, and a means for wiring these plugins together to build a useful system. A few example output plugins have been developed to generate data from the GPS, GSM radio, Wifi radio, accelerometer, magnetometer, and temperature sensors. Additionally, we have developed input plugins for logging the data, uploading the data to a web service, classifying the accelerometer data to detect primitive activities, and clustering locations that a user visits based on wifi fingerprints. The code and applications, and the documentation (which is currently being updated) can be found here. The code is released under the MIT software license, and is thus free to use without warranty, provided that proper attribution is included.
My masters work focused on incorporating variance reduction techniques from the stochastic simulation community into reinforcement learning algorithms in order to improve performance. The rare events paper in the previous section is one example of such a method, and uses adaptive importance sampling as a variance reduction technique.
In Dec. 2007, I took part in the Hierarchical Organization of Behavior: Computational, Psychological and Neural Perspectives workshop at NIPS. The website for this workshop can be found here. I presented some recent work on using Recognizers for learning about options (ie. temporally extended courses of action) in environments with large state and action spaces, where function approximation is employed for state and action representation. (Extended abstract, Slides)
Here are the projects that I completed during my Master's degree at McGill University.
COMP-652: Machine Learning
Final Project on Recognizers in Temporal-Difference Learning with Function Approximation.
ECSE-626: Statistical Computer Vision
Final Project and presentation on creating Illumination Invariant Images using Entropy Minimization.
COMP-627: Theoretical Foundations of Programming Languages
Final Project on Domain Theory and some Partial Orders on the Classical States.
I was on the technical committee for the RL-Competition at ICML'08, and I was the web chair for the 2009 RL Competition. It's a great competition, and a terrific way to try out those fancy Reinforcement Learning algorithms in some fairly tough domains.
My ResearcherID.com badge:
I taught COMP-202: Introduction to Computing 1 at McGill University in the Summer 2007 semester.
I also was a Teaching Assistant at McGill for COMP-202 in Fall 2006 and Winter 2009, COMP-652 (Machine Learning) in Fall 2008, 2009, and 2010, and COMP-102 (Computers and Computing) in Winter 2011.
Additionally, I was the delegate for the School of Computer Science at McGill to the Teaching Assistant union, AGSEM, from 2008-2010.
I have also contributed articles to a couple of online developer websites: