Progress Report

May 16^th, 2006

Overview

TMI re-submission deadline is approaching
Normality Tests
Localised Registration Errors
Noise and Sensitivity
Implementation of another groupwise registration algorithm - awaiting output to assess
Promising improvements to code and re-organisation of past experiments

IEEE TMI Paper Progress

Sent the very latest LaTeX file of the TMI paper
Contains very minor corrections (e.g. typos)
Need to begin re-writing the text fairly soon
Linking of April-May experiments, in reference to reviewers' comments

Localised Registration Errors: Euclidean

Euclidean localised errors, as in the previous report with the shuffle distance
Euclidean on the left hand side
Shuffle in on the right hand side, for comparison

Localised NRR Errors: Euclidean and Shuffle

Sensitivity and Noise in NRR

Redefine ratio for sensitivity measures
See how error bars expand as noise is added and how much noise can be afforded while retaining good sensitivity
Check how sensitivity error affects noise for shuffle and Euclidean distances
Appended are a variety of results with some captions and suitable annotation

Sensitivity and Noise in NRR - Ctd.

Among these figures:
- The effect of noise on Generalisation and Specificity in the type of curves that we have become accustomed to seeing
- The sensitivity of Generalisation and Specificity in isolation
- The effect of noise on the sensitivity of Generalisation and Specificity
- In the latter case, might find it rather odd, inconsistent, and non-patterned

Simplified Generalisation Plots

generalization-euclidean-and-shuffle-5x5-with-noise.png

Simplified Specificity Plots

Sensitivity of the Specificity

sensitivity-of-specificity-euclidean-and-shuffle-5x5.png

Sensitivity of Generalisation

sensitivity-of-generalisation-euclidean-and-shuffle-5x5.png

Sensitivity of the Specificity with Errors

sensitivity-of-specificity-euclidean-and-shuffle-5x5-with-noisified-equivalents.png

Sensitivity of Generalisation with Errors

sensitivity-of-generalisation-euclidean-and-shuffle-5x5-with-noisified-equivalents.png

Normality Test - Context

We have a a histogram of edge lengths for various mockup distributions
We do not know the expected variance
The histogram seems somewhat like a Gaussian when the distributions are identical
We expect this to be a Chi or Chi-squared distribution
Some background information from nist.gov

Notes on Normality test

Chi and Chi-squared distribution explained in various sources, including Mathworld
The Chi normality test - ratio of actual/expected occupancy in histogram bins (liaising with Carole)
Implementation depends on unavailable module (Maple, closed-source MEX files or the Symbolic Toolkit)
The implementation checks Chi-squared goodness of fit for random digits or a roughly normal distribution

Notes on Normality Test - Ctd.

function that could perhaps be used is the following: GoodnessOfFit4RandDigs.m - Performs an alpha significance level chi-square goodness of fit test of randomness on input vector of digits. Makes use of the chiSquarePercentilesBisect and chiSquareProb functions as well as the Matlab Symbolic Toolkit.
Function and toolkit location