Progress Report
January 3
rd
, 2006
Overview
Evaluation based on the entropic graphs method
Dicussion - method in theory and in practice
Setsets and randomisation in entropy estimation
Paper draft progress
References management and 'housekeeping' chores
Approaching talk, deadlines
Entropic Graphs-based Evaluation
An error was corrected, which previously led to badly-scaled results
Cause: brackets were closed at the wrong place
The new curves look more or less the same as before
Further work was done on isolating free parameters (e.g. alpha, constant) so that their effect can be studied
The evaluation framework is much more modular
Interfaces make experiment setup/arrangement trivial and invocation immediate
Various Steps to Follow
No computer resources available throughout the holiday
Remote access is still needed for brute-force-dependent studies
Learn the effect of various parameters by repeating the experiments with different values
Consider alterring:
Alpha value
Constant values
k
as in kNN
Discuss error estimation propagation
Coming Slides that Contain Figures
Shown are some figures which we previously looked at
The error in the entropy equation was eradicated
Attempts were made to discover why we had gotten such high numbers
Ultimately this led to the correct graphs (shown at the end)
Entropy Graphs #1
Entropy of Synthetics->Synthetics and Training->Synthetics
Entropy Graphs #2
Corrected figure (Y-scale changes)
Entropy Versus Generalisation and Specificity
Still need to add the error bars to entropy
Propagation of errors seems complex
Getting more instantiations included in the analysis
Curve are expected to be monotonic given more instantiations
Data which was used for the ISBI submission seems to support this expectation
Entropic Measures Overlaid atop Gen. and Spec.
The curves shown beforehand and their relationship to Generalisation and Specificity (instantiation 1 only)
Survey of Methods
When reversed, one can see how well-correlated entropy is to Gen. and Spec.
These are the corrected plot, which show the entropy as a function of misregistration
The analysis of the numbers is innately the same
Normalisation included, which is a pro
More flexibility available in this form of MST-based calculation, e.g. generalisation to kNN with weighting (see later).
Entropy Graphs #3
Entropic graph-based measures - revised, shown in a clearer way
Entropy Graphs #4
Training set to synthetic set - entropy shown in isolation
Improvements to the Assessment Framework
Added two new options to the model/NRR evaluation framework
Both option are easily controlled from outside the core of the program
Option 1
: Can make the sample of images for distance estimation random
Option 2
: Can use just a subset of a given pool of image rather than iterating over entire sets
Clusters at the School of Computer Science down throughout the holiday
Inability to run large experiments and produce useful results (yet)
Improved Entropy Measures
Implemented additive power weighted kNN
Rather than just minimal spanning trees (MST), consider multiple near nodes
Weighting halves at each stage so if nearest neighbour has weight
x
, the second nearest neighbour has weight
x/2
Might lead to more robust results, but need to run experiments first
Towards a Draft/Submission
Continued work on this draft, which is somewhat presentable
The last few parts are still unorganised
Getting close to having a complete draft which spans roughly 10 pages
Need to encourage colleagues to glance over it, giving some feedback and suggestion
The latest version of the document -
Draft (PDF)
Submission Length Restrictions
Still unable to find information on the number of pages permitted
The instructions are a collection of assorted out-of-date documents and they are not consistent
Submission deadline is February 1
st
Corresponding author yet unknown, but S. Warfield was contacted already
References Management
Migration, assemblage and conversions pending
Management of references in a Web-based PHP/MySQL tool
Somewhat of a free CMS that gained popularity
Open Source, multiple users, categories/groups, collaborative management
Rich graphical interface, requires authentication, available from any Web browser, anywhere in the world
Talk on NRR Assessment
With and without ground truth
Abstract similar to the one embedded in the ISBI submission, so material/content ought to remain similar
Main advancement is entropic graphs
Other advancements less significant and possibly beyond scope (~15 minutes only)
Students Talk on January 27
th
, absence is
accidental
Summary
Framework for evaluating graph entropy
Experiments support the hypothesis
Paper draft in progress
Student Talks (or Monday Meeting) coming up
References management
Paper draft in progress