Kiri L. Wagstaff

My background is in Computer Science, Planetary Science, and Geology. I am most interested in problems that lie at the interfaces between these fields, such as automated methods (artificial intelligence, machine learning) to investigate science questions using planetary data (orbital and in situ).

I am a researcher at the Jet Propulsion Laboratory in Pasadena, CA. After receiving my Ph.D. in Computer Science from Cornell University in 2002, I worked for a year in the JHU Applied Physics Lab's Space Department. My two major projects dealt with space weather prediction and the fault protection system for the MESSENGER spacecraft. I am now a member of JPL's Machine Learning and Instrument Autonomy Group, investigating ways that machine learning can be used to increase the autonomy of space missions. I also teach classes at Cal State L.A. in the Computer Science Department. My projects at JPL have included:

Efficient machine learning to detect transient radio phenomena (e.g., pulsars) in real time
Collaborative machine learning for sensor networks
Automatic landmark identification and change detection in Mars orbital images (dark slope streaks, dust devil tracks, etc.)
Analyzing the sensitivity of machine learning algorithms to high-radiation environments
Predicting county-level crop yield from Earth orbital images
Modeling user preferences for sets, rather than individual items (like music playlists or rover image downlink sets)
Modeling flight software with state charts and using automatic code generation to convert them into C/C++ (for implementation) or Promela (for model checking)
Tracking the north polar ice caps (water and CO₂) on Mars

News and upcoming events:

Newly published:
- Detecting Structural Microbial Biosignatures in Digital Images.
  Kiri L. Wagstaff and Frank A. Corsetti.
  Astrobiology, 10(4), p. 363-379, doi:10.1089/ast.2008.0301, May 2010.
  - Given an image of a rock that contains layered structures, is it possible to determine whether the layers were created by life (biogenic)? We evaluated several quantitative measures that capture the degree of complexity in visible structures, in terms of compressibility (to detect order) and the entropy (spread) of their intensity distributions. None of the techniques provided a consistent, statistically significant distinction between all biogenic and abiogenic samples, but the PNG compression ratio provided the strongest distinction and could inform future techniques.
- Confidence-Based Feature Acquisition to Minimize Training and Test Costs (pdf, 11 pages, 624K).
  Marie desJardins, James MacGlashan, and Kiri L. Wagstaff.
  Proceedings of the SIAM Conference on Data Mining, p. 514-524, April 2010.
  - How can the best classifier be learned, when data features may be missing both while training and when classifying new items? Which missing features should be acquired, and in what order? This paper presents a greedy approach to achieving good performance with low feature acquisition costs.
Current activities:
- I am an organizer of EAAI-10: The First Symposium on Educational Advances in Artificial Intelligence (July 13-14, 2010)
- I am also helping to organize CIDU-10: Conference on Intelligent Data Understanding (October 5-7, 2010)
I served as part of Crew 89 at the Mars Desert Research Station, from January 22 to February 7, 2010. You can read our blog to read out what we did, see pictures, and watch videos!
I was selected to receive the 2008 Lew Allen Award for Excellence in Research for "advancing the performance and application of machine learning methods to onboard Earth science missions and spacecraft engineering."