Paul K. Mazaika, PhD

Physical Science Research Associate

Center for Interdisciplinary Brain Sciences Research

Psychiatry and Behavioral Sciences, Stanford University

 

Research

How does functional activity in the brain reflect the difference between healthy and mentally impaired populations? Our first objective is to create new methods for accurate fMRI analysis of clinical data to better characterize these populations. At a deeper level, a functional map represents a "footprint" of underlying neural activities. These activites are computationally complex for even the most simple behaviors. How are the multi-tasking and adaptive abilities of people supported by the distributed neural architecture, and how does experience combine with genetic and neurochemical factors to adapt neural networks into successful behaviors? Our second objective is to understand how development influences information processing and learning to illuminate how these factors impact human behavior.

C.V.

Publications

Experience

FMRI Software

 

_________________________________________________________________________________________________________________________

Curriculum Vitae


Academic Experience

2006 – present Physical Science Research Associate, CIBSR, Stanford University

2004 – 2005 Physical Science Research Assoc., Gabrieli Cognitive Neuroscience Lab, Stanford University



Industry Experience

2002 – 2003 Imaging Engineer, Foveon, Inc.
Developed automated quality tests of imaging chips for Sigma digital cameras

1996 – 2002 Manager, Verification and Computer Aided Design,
Developed imaging chips used in Xerox DC 212/214 copiers

1992-1996 Consultant, Hughes Electro-Optical Systems, and Lockheed Systems
Provided expertise on real-time adaptive image processing techniques

1989 – 1992 Test Director, Systems Engineer, Lockheed Space Systems
Managed software development for multi-modal image data fusion applications

1986 – 1989 Research Technical Staff, Northrop Research and Technology Center
Developed adaptive 3-D numerical grid algorithms for parallel computers

1983 – 1986 Manager, Computer Science Research Lab, Aerospace Corporation
Developed artificial intelligence methods and software for real-time decision support

1978 – 1983 Manager, Systems Engineering, Aerospace Corporation
Optimized systems design for the Global Positioning System (GPS)

1974 – 1978 Satellite Systems Engineer, Aerojet Electro-Systems Company
Optimized sensor and signal processing design for infrared satellites.


Honors

2006 NIMH Grant K25MH077309. Improved Methods for Single Subject FMRI Analysis for Clinical Application.

1997 Xerox Team Excellence Award

1992 Lockheed BSTS Team Accomplishment Award

1987 Defense Support Program (DSP) Program Recognition Award

1986 Aerospace President’s Award for Analytical Achievement
“For developing analytical techniques which resulted in the discovery that a
non-uniform18-satellite Global Positioning (GPS) Constellation satisfies the same
objectives as the original 24-satellite Constellation, with large cost savings to the
Program.”

1986 Aerospace Engineering Group Award

1970 Mathematics Medal – New York University

1970 Physics Medal – New York University

1970 NSF Fellowship


Education

Ph.D. 1974, Applied Mathematics, Caltech. Advisor: P.G. Saffman
Thesis: On the Settling Speed of a Dilute Array of Spheres


B.S. 1970, Applied Mathematics, New York University

Click to Return to Top

___________________________________________________________________________________________________________________

Publications


fMRI Publications and Presentations

Mazaika, P.K., Whitfield-Gabrieli, S., and Reiss, A.L., Artifact Repair and Validation of Estimates from fMRI Data from High Motion Clinical Subjects, (submitted)


Mazaika, P.K., Whitfield-Gabrieli, S., Reiss, A.L., Artifact Repair for fMRI Data from High Motion Clinical Subjects, invited presentation at Human Brain Mapping, 2007


Hoeft, F., Ueno, T., Reiss, A.L., Meyler, A., Whitfield-Gabrieli, S., Glover, G., Keller, T.A., Kobayashi, N., Mazaika, P., Jo, B., Just, M.A., and Gabrieli, J.D.E. Prediction of Children's Reading Skills Using Behavioral, Functional And Structural Neuroimaging Measures. Behavioral Neuroscience, (in press).


Mazaika,P.K., Golarai, G., and Gabrieli, J.D.E., Classifying Single Trial fMRI: What can machine learning learn?, Presentation at NIPS 2006 Workshop on Decoding Mental States, Dec. 9, 2006.


Mazaika, P.K., Whitfield, S., Cooper, J.C., Detection and Repair of Transient Artifacts in fMRI Data, NeuroImage 26, Supplement 1, S36. HBM 2005.


Mazaika, P.K., FMRI Dynamics: from Artifacts to Activations, Neuroscience Institute at Stanford fMRI Colloquium, March 2005


Mazaika, P.K., Survey of Single Voxel White Noise, Stanford Technical Report, Nov.12, 2004.



Other Publications

O’Malley, R.E. Jr., Mazaika, P.K., On the Asymptotic Solution of Multi-Point Boundary Value Problems with Discontinuous Coefficients, Indiana University Mathematics Journal, Vol. 20, No. 7, 1971

Book, S.A., Brady, W.F., Mazaika, P.K., The Non-Uniform GPS Constellation, IEEE PLANS 80, Position Location and Navigation Symposium, Dec. 1980.


Mazaika, P.K., Orientation of Measurement Sensors for Optimum End-of-Life Performance. IEEE Transactions on Aerospace and Electronic Systems, Vol. AES-17. No. 2, March 1981.


Mazaika, P.K., Jitter Induced Clutter in Staring Sensors Arising from Background Spatial Radiance Gradients. Proc. SPIE Conf. Modern Utilization of Infrared Technology VII, Aug. 1981.


Mazaika, P.K., Sarrafian, H., Thompson, T., Analysis of Background Spatial Radiance Variations for RM-19 and CAMP Data. Proc. IRIS Specialty Group on Targets, Backgrounds, and Discrimination, Dec. 1981.


Bergen, T.L., Mazaika, P.K., Evaluation of Spatial Filters for Background Suppression in IR Mosaic Sensor Systems. Proc. SPIE Conf. On Real Time Signal Processing, May 1982.


Mazaika, P.K., Statistical Analysis of Background for Remote Sensing Applications. Methods of Nonlinear Analysis Session, SIAM National Meeting, July 1982.


Mazaika, P.K., Starer vs Scanner: A Background Suppression Comparison. Proc. SPIE Conf. On Focal Plane Methodologies III, Aug. 1982.


Mazaika, P.K.., Starer vs Scanner: A Background Suppression Comparison. Proc. DARPA Infrared Sensor Signal Processing Workshop, Oct. 1982


Mazaika, P. K., Jitter Induced Clutter in Staring Sensors Arising from Background Spatial Radiance Gradients. Optical Engineering, Sep/Oct 1982.


Mazaika, P.K., A Local Spatial Correlation Length for Background Modeling. Proc. of the International Society for Optical Engineering (SPIE) Conference On Infrared Technology IX, Aug. 1983, pp. 7-23.


Mazaika, P.K., A Lattice Summation Using the Mean Value Theorem for Harmonic Functions. SIAM Review, Vol. 26, No. 1, Jan. 1984.


Landauer, C., Feldman, P., Mazaika, P.K., Fault Tolerance from Collective Behavior. Fault Tolerant Computing Symposium, IEEE Computer Society, June 1985.


Sweet, M.M., Mazaika, P.K., An Heuristic Algorithm for a Class of Integer Programming Problems. SIAM Symposium on the Complexity of Approximately Solved Problems, April 1985.


Gillam, A., Mazaika, P.K., BIRDMAN: An Expert System for False Event Elimination. Proceedings of Conference of Military Applications of Electro-Optical Technology, April 1986.


Mazaika, P.K., Gillam, A., High Performance Discrimination by Modeling an Expert. Proceedings of DARPA Strategic Systems Symposium XII, Oct. 1986.


Mazaika, P.K., A Lattice Summation Using the Mean Value Theorem for Harmonic Functions. Mathematical Modeling, Society of Industrial and Applied Mathematics (SIAM) Publications, 1987.


Mazaika, P.K., A Mathematical Model of the Boltzmann Machine. Institute of Electrical and Electronics Engineers (IEEE) First International Conference of Neural Networks, Vol. III, June 1987, pp. 157-163.


Mazaika, P.K., Implementation and Technology Transfer of an Expert System in an Aerospace Company. American Association of Artificial Intelligence (AAAI) Workshop: Stanford, July 1992.


Click to Return to Top
____________________________________________________________________________________________________________________

Experience

A few sample projects from industry….

Algorithms to imitate the human inspection process

Digital imaging chips must provide flawless pictures, but many chips have small defects due to the complexity of manufacturing millions of nanometer-sized circuits. Automated quality testing of the chips is necessary to select good chips from bad chips. We modeled the visual perception sensitivities of professional photographers, in order to establish test methods and accuracy thresholds for the automated assembly line. The software is currently applied to Foveon digital sensors for use in Sigma and Polaroid cameras.

Design of a distributed image processing system

Good color copiers discriminate text from images, then make the text crisp and black, and the images accurate and smooth, in order to produce high quality copies. High speeds are accomplished using parallel and pipelined chip designs. The multiple processors on each chip must execute different algorithms, and communicate to each other and to shared memory, in near real-time without any errors. As part of the design team, I designed imaging algorithms and managed the software automation necessary for chip design and verification. The chips were used in Xerox DC 212/214 copiers.

Neural network research

Hopfield neural networks were considered as an approach for highly fault-tolerant computer processing. We measured the rates of convergence to the attractor states, and developed unlearning algorithms to prevent being trapped in suboptimal states. When additional noise in the networks changes attractors into quasi-stable states, we calculated expected durations for different states, and showed that networks could be run asynchronously with the same pattern completion properties. While these networks had some biological analogies, they were not as hardware efficient as other approaches to fault-tolerance. More robust functionality was obtained using strong Hamming error-correcting codes for memory and redundant processors that voted on the correct answers.

Artificial intelligence and cognitive modeling

Trained human operators are used to monitor satellite imagery. While computers can perform detection, estimation, and pattern classification, people are better at assessing the overall validity of the data. This research attempted to cognitively model the operators in order to automatically “understand” the contextual properties in the images. We developed a multi-path pattern recognition and expert system approach, added robustness to input variations and data anomalies, and included a self-assessment of the computer system itself. The high performance of the demonstration project led to some unanticipated engineering problems. Testing was difficult because of its very high accuracy rate, and "trust" in the algorithms was difficult because of its multiple logical paths. An unexpected issue was human factors - a very accurate system would make the humans-in-the-loop potentially less effective due to lack of activity (and associated boredom). These issues remain open research questions, and the project was not implemented.

Design optimizations for Global Positioning System

The Global Positioning System (GPS) provides accurate navigation for users around the world. Performance depends on the viewing angles to the satellites for each user, while the satellites are continually moving in their orbits. A simulation of satellite geometries for all locations on the earth, for all times in the orbits, was required to estimate worldwide navigation performance. We developed mathematical methods to evaluate hundreds of different constellations in a short amount of time, which enabled fast optimization of the constellation design. GPS also provides extremely accurate time (tens of nanoseconds), because each nanosecond of time error causes about one foot of navigation error. Time errors are caused by clock instabilities, special relativity effects, and the reduction of the speed of light in the atmosphere. We specified thresholds of atomic clock accuracy and stability, and the number and locations of ground stations needed to update the clocks, in order to provide sufficiently accurate time measures to meet navigation objectives.

Click to Return to Top

______________________________________________________________________________________________________________________

Software for fMRI


Artifact Repair for fMRI data from Clinical Subjects

Thousands of fMRI data sets have been collected for many types of patients, but the data is difficult to analyze because clinical subjects often move during a scan session. The artifact repair algorithms improve the analysis of clinical data sets by detecting outliers in the data caused by subject motion or transient noise, and then repairing the bad data by replacing outliers with more reasonable interpolated values. The accuracy of the repairs can be visually evaluated from movies of the signal variation of data from all the voxels, and by reviewing the Global Quality measure of the statistical efficiency of the GLM estimates.

Software is available at: http://cibsr.stanford.edu/tools/ArtRepair/ArtRepair.htm.


Pattern Classification for Single Trial fMRI Data

Pattern recognition algorithms and software were developed to detect and classify responses to single subject, single trial fMRI data. Special filtering algorithms suppress physiological fluctuations to improve classification, and allow BOLD activations from a single trial in an event-related experiment to be displayed. A sample image of a single trial is shown below. The classification algorithms "read" the fMRI images and correctly identify Face vs. Non-Face input on 90% of the trials. The software was first used to study cognitive development of the face area in children by G. Golarai. Single trial classification accuracy is one measure of repeatability of neural activations for single subjects. The visual repeatability of single trials allows detailed checks on data quality, and identifies large outliers to discard from noisy data sets- with the significant implication that artifact repair could “rescue” previously unusable clinical fMRI data sets by discarding bad data.



Figure: Example of SINGLE SUBJECT, SINGLE TRIAL BOLD Response on fMRI slices through fusiform gyrus and amygdala. Face response (Left) and Object response (Right). Data courtesy of G. Golarai. Images were obtained from the contrast movie of the ArtRepair software.



Click to Return to Top

Last Updated: March 2007

Copyright 2007: Paul Mazaika