Difference between revisions of "Sean G. Carver's Research Interests"

From Sean_Carver
Jump to: navigation, search
(→‎Pedagogy for Training Students on System Identification)
 
(39 intermediate revisions by the same user not shown)
Line 1: Line 1:
−
'''Broadly, my interests involve system identification (SysID) applied to biological systems,''' especially:
+
'''Much of my current research involves projects related to the statistical analysis of models using simulated data:'''
−
 
 
−
* developing methods of SyID,
 
−
* applying methods of SysID to understand biological processes, and
 
−
* advancing methods for teaching SysID (pedagogy).
 
  
−
By ''system identification'', I mean ''using statistics, derived from experimental data, to constrain dynamical models of a system''.  Typically, SysID involves parameter estimation, model selection, and model validation.
+
* '''Overlapping software projects:''' [https://github.com/seancarverphd/klir KLI-R] (R/Github/Git) and [https://bitbucket.org/seancarverphd/kli/ KLI] (Python/Bitbucket/Mercurial).  KLI stands for "Kullback-Leibler Interactive."  These projects involve, among other things, computing the number of samples needed to reject an alternative model with the likelihood ratio test, in favor of a true model that produces the data.   
 +
::Future Conference Proceeding: I will present this work at the Joint Statistical Meeting (JSM) in Baltimore, August 2017.
  
−
== System Identification of Cooperative Control ==
+
* '''Baseball:''' how many innings must be played by model of the Baltimore Orioles (fitted from actual Orioles home games) to reject the model that the  New York Yankees are playing?  This statistic provides an interpretable way of quantifying the similarity of models.
 +
::Student Collaborator (just graduated, but still working with me) Rebeca Berger.
 +
::Future Conference Proceeding: I will present this work at the Joint Statistical Meeting (JSM), Baltimore, August 2017.
  
−
When people cooperate to perform a task that requires that they stay in sync (e.g. salsa dancing, or chamber music), how do they maintain synchrony given that their biological clocks are noisy? I am currently involved in designing experiments (in collaboration with the LIMBS lab, Johns Hopkins) to collect and analyze data from humans performing simple sensorimotor tasks where they must cooperate to stay in sync.  We plan to analyze the data with SysID. Please see [[Media:carverACC.pdf|a preprint of a relevant paper (click here)]].
+
* '''Motor Control:'''  With my student collaborators, I have been looking at continuation tapping, an experimental paradigm involving a metronome, and subject tapping to the beat.  After a sufficient duration of time, the metronome stops, and the subject must keep the same rhythm.  In an effort to perform system identification of the internal clock for motor control, we found that the inter-tap intervals are fit equally well by the Normal and the Inverse Gaussian distributions, and both fit much better than the Laplace distribution.  Contrary to this finding, the only relevant study in the literature we discovered, a review paper, reported that inter-tap intervals have a Laplace distribution.  This review paper did not provide many details about how the data were collected and analyzed. I am working with Daniel Scanlan to explore, through simulations, when models of continuation tapping produce data that fit the Inverse Gaussian/Normal distributions (these two are almost identical at our parameters) and when they produce data that fit the Laplacian distribution.
 +
::Student Collaborators: (current) Daniel Scanlan, (former) Wasim Ashshowaf, and (former) Alexander Spinos
 +
::Future Paper: Daniel and I plan to submit this work for publication, probably in PLoS One.
  
−
== Control Theory Where Agent(s) Rely on Noisy Clocks ==
+
*'''Ion Channels in Neuroscience:'''  Ion channels provide much of the molecular basis for neural signaling. Models of ion channels are continuous time Markov chains with hidden states, far more complicated than any of the applications above.  Much of the KLI in python code involves these models of ion channels.
  
−
No clock tells time perfectly [http://www.nist.gov/pml/div688/logicclock_020410.cfm].  In the past, when timekeeping was not good enough for a particular engineering application, the engineers invariably solved the problem by building better clocks.  In solving the problem this way, they avoided needing to create better control algorithms to deal with noisy clocks.  Invariably, they managed to get away with assuming perfection in timekeeping.  Unfortunately, evolution has different constraints, and it is apparent that to understand the sensorimotor control of behaviors such as dance, we need to extend control theory to deal with the case that internal clocks are noisy.  To date, surprisingly little has been done in this regard, but the models I pose for the SysID of cooperative synchrony do involve noisy internal clocks.
+
'''Other projects:'''
  
−
== Pedagogy for Training Students on System Identification ==
+
* '''Analysis of data from the whole brain larval zebrafish at cellular resolution.'''  A collaborator from Janelia Research Institute, [https://www.janelia.org/people/misha-ahrens Misha Ahrens] provided me with a data set that consisted of about 100,000 time series, one from each of 90% of the neurons in a larval zebrafish. The time series consisted of measurement of calcium from within their respective cells, using a calcium sensitive dye. The fish were involved in a closed-loop sensorimotor behavior, similar to swimming "like Neo in the matrix."  Specifically the fish were actually immobilize in a microscope, but were given a visual stimulus, as if they were swimming.  The harder their nervous system commanded the tail to wag, the faster the stimulus moved, in the appropriate direction.  Each time series had the same length, about 4000 samples.  I looked at the intrinsic dimension of the 4000 points embedded in 100,000 dimensions.  I found that the intrinsic dimension was about 15.  I plan to see if this result is consistent from fish to fish, and also look at the persistent cohomology of the data.
  
−
I am interested in finding effective methods for training students in the art of system identification.   Training students to use SysID is easier than might first be thought, because one can collect the needed experimental data from ''simulations''.   For training and testing purposes, using a simulated system to generate the data is desired, because the experimenter then knows the correct model ahead of time.  Knowing the correct model allows an assessment of the suitability and effectiveness of the SysID methods.  Moreover, an instructor can create a dataset from simulation, but withhold from students the structure and parameterization of the generating model, giving students the more realistic experience of not already knowing the answer.  In my mind, the primary advantage of this feature comes from the fact that no large investment in a laboratory equipment and training is needed to train students in this advanced laboratory technique.
+
* '''Server administration, data collection, data storage, and data analysis:''' I set up a [http://stat370.com web server] in my home to be used by my students in a recent statistics class I taught.  I had also wanted to let my students run background processes collecting data from the web for analysis (either through APIs or through scraping and crawling the web). It became clear that the best way for students to collect data in this way is to have them rent their own server from Amazon Web Services. This summer I am going to work with Jennifer Schaffer, a student from the class, to collect, archive, then analyze a large volume of tweets concerning the U.S. politics.  We plan to use MongoDB to store the tweets, but we have not determined how they will be analyzed.  We also plan to look at Twitter's follower network, using a Neo4j database.  Finally, I plan to migrate the web server from my own machine to a (separate) Amazon Web Services host, and keep it active for students, past, present, and future, to use.
 +
::Student Collaborator (just graduated, but still working with me): Jennifer Schaffer
 +
 
 +
* '''Cashflow:''' As a fun project to learn more about SQL and Regular Expressions, I set up my computer to ingest data from my bank and credit card company.  I plan to archive it in a database, and provide regular reports.  The code is presently in a private repository.

Latest revision as of 03:09, 14 May 2017

Much of my current research involves projects related to the statistical analysis of models using simulated data:

  • Overlapping software projects: KLI-R (R/Github/Git) and KLI (Python/Bitbucket/Mercurial). KLI stands for "Kullback-Leibler Interactive." These projects involve, among other things, computing the number of samples needed to reject an alternative model with the likelihood ratio test, in favor of a true model that produces the data.
Future Conference Proceeding: I will present this work at the Joint Statistical Meeting (JSM) in Baltimore, August 2017.
  • Baseball: how many innings must be played by model of the Baltimore Orioles (fitted from actual Orioles home games) to reject the model that the New York Yankees are playing? This statistic provides an interpretable way of quantifying the similarity of models.
Student Collaborator (just graduated, but still working with me) Rebeca Berger.
Future Conference Proceeding: I will present this work at the Joint Statistical Meeting (JSM), Baltimore, August 2017.
  • Motor Control: With my student collaborators, I have been looking at continuation tapping, an experimental paradigm involving a metronome, and subject tapping to the beat. After a sufficient duration of time, the metronome stops, and the subject must keep the same rhythm. In an effort to perform system identification of the internal clock for motor control, we found that the inter-tap intervals are fit equally well by the Normal and the Inverse Gaussian distributions, and both fit much better than the Laplace distribution. Contrary to this finding, the only relevant study in the literature we discovered, a review paper, reported that inter-tap intervals have a Laplace distribution. This review paper did not provide many details about how the data were collected and analyzed. I am working with Daniel Scanlan to explore, through simulations, when models of continuation tapping produce data that fit the Inverse Gaussian/Normal distributions (these two are almost identical at our parameters) and when they produce data that fit the Laplacian distribution.
Student Collaborators: (current) Daniel Scanlan, (former) Wasim Ashshowaf, and (former) Alexander Spinos
Future Paper: Daniel and I plan to submit this work for publication, probably in PLoS One.
  • Ion Channels in Neuroscience: Ion channels provide much of the molecular basis for neural signaling. Models of ion channels are continuous time Markov chains with hidden states, far more complicated than any of the applications above. Much of the KLI in python code involves these models of ion channels.

Other projects:

  • Analysis of data from the whole brain larval zebrafish at cellular resolution. A collaborator from Janelia Research Institute, Misha Ahrens provided me with a data set that consisted of about 100,000 time series, one from each of 90% of the neurons in a larval zebrafish. The time series consisted of measurement of calcium from within their respective cells, using a calcium sensitive dye. The fish were involved in a closed-loop sensorimotor behavior, similar to swimming "like Neo in the matrix." Specifically the fish were actually immobilize in a microscope, but were given a visual stimulus, as if they were swimming. The harder their nervous system commanded the tail to wag, the faster the stimulus moved, in the appropriate direction. Each time series had the same length, about 4000 samples. I looked at the intrinsic dimension of the 4000 points embedded in 100,000 dimensions. I found that the intrinsic dimension was about 15. I plan to see if this result is consistent from fish to fish, and also look at the persistent cohomology of the data.
  • Server administration, data collection, data storage, and data analysis: I set up a web server in my home to be used by my students in a recent statistics class I taught. I had also wanted to let my students run background processes collecting data from the web for analysis (either through APIs or through scraping and crawling the web). It became clear that the best way for students to collect data in this way is to have them rent their own server from Amazon Web Services. This summer I am going to work with Jennifer Schaffer, a student from the class, to collect, archive, then analyze a large volume of tweets concerning the U.S. politics. We plan to use MongoDB to store the tweets, but we have not determined how they will be analyzed. We also plan to look at Twitter's follower network, using a Neo4j database. Finally, I plan to migrate the web server from my own machine to a (separate) Amazon Web Services host, and keep it active for students, past, present, and future, to use.
Student Collaborator (just graduated, but still working with me): Jennifer Schaffer
  • Cashflow: As a fun project to learn more about SQL and Regular Expressions, I set up my computer to ingest data from my bank and credit card company. I plan to archive it in a database, and provide regular reports. The code is presently in a private repository.
Retrieved from "https://seancarver.org/index.php?title=Sean_G._Carver%27s_Research_Interests&oldid=2893"