Meeting Notes 20101006 NEMO core meeting - ERP pattern mapping

From Nemo

Jump to: navigation, search
  • Convened by: NEMO core
  • Presenter: Haishan Liu
  • Notes by: Gwen Frishkoff
  • Where: WebEx
  • Attendees: Gwen, Allen, Dejing, Haishan, Kurt, Jason, Bob, Paea, Snezana



The goal for this meeting was for Haishan to give an update on his work extending the PAKDD'10 work on metric mapping to develop methods for mapping ERP patterns across different datasets.


To download power point slides, please click on "Presentation_20101006_WebEx_ERP_Pattern_Mapping_-_Mapping_across_ERP_Patterns_from_Simulated_Data.pptx" (from link at [[1]]).

Gwen's notes:

  • simulated ERP data
    • PCA processed & clustered
    • ICA processed & clustered
  • matching process (blind, automated)
    • labeled data (e.g., MFN =cluster2+9)
    • changed meaningful labels to 'Cluster 1', 'Cluster 2'
  • So Haishan's method appears to be characterizing (post-hoc) the attributes that belong to expert-labeld data
  • Method = density profile
    • appropriate for mapping clusters across 2 dataset that do not share any observations in common
    • need to define density of "attribute-bin region" for each cluster (how many observations in a "region" belong to cluster)??
    • density profile = characterizes how attributes are distributed for each cluster
    • bin = level of attribute
      • e.g., Vc = density profile on first level of an attribute for each cluster
  • Then similarity defined as product of density profiles (?)
  • But there are multiple products that can be computed (?).
  • Optimization?
    • Select largest number resulting from each product.
    • Can also look at Euclidean distance between clusters as validation (?).
  • We discussed whether this method is robust to differences in size/scale of different clusters. Need to make sure that product doesn't give inaccurate results when there is a difference in length of two vectors.
  • Hungarian method for minimizing diff between two vectors (i.e., mapping between clusters) -- using TImax (peak latency) & means for different spatial ROI.
  • Assumptions:
    • Allen: Does this assume 1to1 mapping between patterns from two datasets? Yes
    • Gwen: can this be relaxed? Maybe (probably not)
  • Look back at attribute stats (means for spatial, temporal attributes) to interpret full set of results from Hungarian method.
  • Extend this to "raw" clustering results (before expert post-processing). Need to constrain number of clusters to be the same (unrestricted clustering; just constrain selection).
  • 100% accuracy (correspondence with gold standard)
  • ADCO numbers give sense for robustness of each set of clustering results.

Action items

  • Conference paper due October 15. Haishan will send draft to Dejing, Gwen, & Bob this weekend
  • Gwen to work with Haishan on substantive interpretation of results (e.g., ADCO scores)
  • Dejing suggested that we extend conference paper to produce Neurocomputing (journal) paper submission for December

WebEx Recorded Archive

If you wish to review the recording of the web conference, click the link below:


NEMO Teleconference-20101006 1802-1 October 6, 2010, 3:29 pm New York Time 1 hour 17 mins

/wiki/images/1/17/Fish1.png /wiki/images/e/ea/Fish2.png /wiki/images/f/fa/Fish3.png /wiki/images/f/ff/Fish4.png /wiki/images/4/40/Fish5.png /wiki/images/c/c5/Fish6.png
Personal tools