Tin Kam Ho

1
Interactive Pattern Discovery with Large Imaging
Databases

• Tin Kam Ho
• Computing Sciences Research Center
• Bell Labs, Lucent Technologies
• In collaboration with
• David Wittman, J. Anthony Tyson of UC Davis
• Samuel Carliles, William OMullane, Alex Szalay
  of JHU

2
What Is the Story in this Image?
3
Solving the Puzzle with a 3-step Approach

1. Describe each symbol shape with a numerical
   vector 23 12 17 28 11
2. Find clusters of symbol shapes
3. Interpret each cluster using context

4
10.10.10 51.37.50.54.41.35.37 39.47.33.44
13.13 33.52.6.52 83.65.73.68 73.84
72.65.83 83.69.84 65 71.79.65.76 79.70
82.69.83.84.79.82.73.78.71 83.69.82.86.73.67.69
79.78 73.84.83 76.79.78.71.13.68.73.83.84.65.7
8.67.69 78.69.84.87.79.82.75 70.65.83.84.69.82
84.72.65.78 84.72.69 83.89.83.84.69.77 73.84.
83.69.76.70 68.73.83.67.79.78.78.69.67.84.83
67.65.76.76.83 65.70.84.69.82 65
67.65.66.76.69 66.82.69.65.75.14
SERVICE GOAL -- ATT said it has set a goal
of restoring service on its long-distance network
faster than the system itself disconnects calls
after a cable break.
5
Tracking Intensive Rain Cells in Radar Images
6
The Deep Lens Survey(Tyson, Wittman, )
BVRz to 26 mag over 28 sq. degree
http//dls.physics.ucdavis.edu/
7
Weak Gravitational LensingUses distortion of
background galaxies to map foreground mass
concentrations
J.A. Tyson, DLS 2002
8
Catalog of Extracted Objects
9
Stars or Galaxies?
J.A. Tyson, DLS 2002
10

• Discrimination task depends on tiny differences
  in color and shape
• Survey is to an unpreceded depth most objects
  have never been observed before and nobody knows
  their true classification
• How does one build confidence on the results of
  the classifier?
• Need to correlate several perspectives object
  characteristics in the color space, shape
  parameters, the brightness statistics
• Visualization can help verify correctness of
  preprocessing steps, clean up undesirable
  artifacts, choose relevant samples, spot
  explicit patterns, select useful features, and
  suggest algorithms and models

11
The Virtual Observatory
http//www.us-vo.org/
http//www.ivoa.net/
12
Essential Steps in Automatic Pattern Recognition
Samples
Supervised learning
Unsupervised learning

features
features
Feature Extraction

Classifier Training
Clustering
classifier
feature 2

Cluster Validation
Classification

Cluster Interpretation
class membership
feature 1
13
Data Relationships Across Multiple Feature Sets
Data Mining
Simulation Analysis
Parameters
Responses
Feature Set A
Set B
Feature Computation
Unknown Relationship
Filtering, Clustering
Clustering
14
Key Algorithms

• Clustering
• find natural groups in data, construct index
  structures to facilitate proximity queries
• Dimensionality reduction
• embed high-dimensional data in 2D displays
• Navigation
• traverse index structures in systematic ways

15
Clustering Methods

• Model based Clustering
• identification of finite mixtures
• Partitional Clustering
• divides data set into N mutually exclusive
  subsets
• Hierarchical Clustering
• top-down procedures tree splitting
• bottom-up, agglomerative procedures merge
  similar clusters successively

16
Similarity / Clustering of Objects from
Different Perspectives

• Objects can be described by many types of
  attributes
• position, weight, shape, spectrum, time
  variability,
• Meaningful similarity metric exists only for the
  same type of attributes
• Clusters found from one perspective need to be
  correlated to those from others
• e.g. Are the objects similar in color also
  similar in shape?

Shape clusters
Color clusters
17
Exploratory Tools Needed

• To bring in domain expertise, interpretation
  context
• To visualize data or classifier geometry
• To track point/class correlations
• To test tentative classifications
• To compare groupings from different perspectives
• To relate numerical data to other data types
• To facilitate systematic, repeatable explorations

18
Mirage for Interactive Pattern Recognition
http//www.cs.bell-labs.com/who/tkh/mirage

• Data Display in Linked Views
• Show patterns in histograms, scatter plots,
  parallel coordinates, tables, and images
• Selection and Tracking
• Select points in any view, broadcast to all
  others
• Traversal of Data Structures
• Walk in histograms, cluster graphs or trees,
  echoed in all other views
• Graphical Utilities
• Open multiple-page plots with arbitrary
  configuration
• Command Scripts
• Run prepared groups of operations as an animation

• Intuitive Graphical Tool for
• Exploratory Data Analysis
• Visualization of Clusters and Classes
• Correlation of Proximity Structures
• Manual or Automatic Classification

19
Software Features

• Based on Java Swing library
• Intuitive, easy-to-use graphical operations
• Mutiple-page, arbitrary plot configurations
• Online or offline cluster analysis
• GUI or Script driven command execution
• Database interface via JDBC
• Ready to be adapted for on-line monitoring
• Ready to be integrated with database access and
  decision support systems

20
Design Motivated by the Needs

• Interactive plays, intuitive operations
• to bring domain experts into the loop
• Multiple types of plots, extensible for more
• to visualize data or classifier geometry
• Linked views, traversal actions
• to track point/class correlations
• Highlights, colors
• to test tentative classifications
• Projection to arbitrary subspaces
• to compare groupings in different
  perspectives
• Linking data with images
• to relate numerical data to other types
• Command scripting
• to facilitate systematic, repeatable
  explorations

21

• Challenges for the Analysis Tool
• Separate treatment of non-comparable groups of
  variables
• Versatile visualization utilities allowing many
  perspectives
• Support for exploratory discovery across diverse
  data types
• Integrate manual automatic pattern recognition
  methods
• Also, a good tool should
• -- leverage existing visualization and analysis
  methods
• - enable continued growth new visualization,
  analysis tools
• - support interface with existing databases
• - be scalable in data volume and processing
  speed

22
Mirage Core
Towards Extensibility
External Rendering Code
VO
Data Archives
Custom Data Views
Data Access Clients
Cone Search, CAS
FITS viewer,
Python? Matlab?
Extinction Calculator
Data Analysis Methods
Data Exchange Pipes
Other Analysis Platforms
Web Services
23
VO Enabled Mirage(with Samuel Carliles, William
OMullane, and Alex Szalay)
24

• VO Enabled Mirage
• http//skyservice.pha.jhu.edu/develop/vo/mirage/
  
• Load VOTable data and perform VO Cone/SIAP and
  
• SDSS CAS searches using IVOA Client Package
• Astronomical imaging module loads FITS images
• using JSky classes, supporting image
  operations
• Select data points and broadcast selection to
  other views. Cut levels. Colormap. SAO
  DS9-style brightness/contrast enhance. Zoom.

25
Extinction Web Service(with Chris Miller,
Simon Krughoff)Using DIRBE/IRAS Dust Maps by
Schlegel et al.
26
More at NVO Public Release 1.0
205th Meeting of the American Astronomical
Society 9-13 January 2005 San Diego, CA
Wednesday, 12 January Astronomical Research with
the Virtual Observatory
27
Analysis of Simulations of Control Dynamics in
Optical Transport Systems(with the FROG
collaboration)
Fiber link
Head End Terminal
Repeater
Repeater
Gain Equalizer
Repeater
Repeater
Signal Spectrum with noise floor
Tail End Terminal
28
Monitoring Network Traffic
(With Marina Thottan, Ken Swanson)

• Software tool for online monitoring and analysis
  of QoS in IP networks
• continuously monitors traffic statistics at edge
  and core devices
• synthesizes statistics in real time to obtain
  network-wide QoS status and general network
  element health indicators
• Mirage refreshes displays on alerts of database
  updates via Java Messaging Service

DiffServ Edge(aggregation andclassification)