Content: developments throughout the period of time which precedes January 20th
Listed chronologically
Used for recapitulation of points last explored
Ends with some dates and key steps ahead
Might need closer involvement from various parties
General Notes on Starting Point
We were dealing with entropy of simple Gaussian distributions
It was obvious there was still something wrong
When the displacement of the dynamic distribution (one remains static) is zero, the two distributions are, by definition the same
The entropy calculation functions were over-customised to deal with the real data
Having found two errors and fixed them, we finally got 'sane' results
Here is the output from some fairly coarse 'toy data'
Entropy of Two Simple Distributions
Larger Experiment: A->A, A->B, Overall Entropy
The curves look much 'cleaner' given a large enough sample
In this particular case:
number_of_examples_static=50; % AKA training set
number_of_examples_deformed=1000; % synthetic set
numbers_of_dimensions=100;
number_of_repetitions=10;
Larger Sample for Entropy Graph
Further Progress
These plots at last looked sensible
Work was done on visualising the distributions
Gamma Variation and the Effect on Entropy Curves
Parameters, as agreed at an earlier stage:
large number of dimensions: 1,000 (10,000 would have taken too long)
points in dynamic distribution (synthetic set): 1,000
points in static distribution (analogous to training set): 50
sample from dynamic distribution (for estimation of distances to self): 50
10 repetitions
10 levels of increasing displacement
Gamma = 1 , 2 , 3 , 4
Gamma Variation - A Few More Notes
Plotted the same entropy curve
Head-to-head, tail-to-tail - eventually plotted on the same figure
What we expect to change: slope, error bars, constant (although subtractions annuls its effect)
Shortly afterward: generated plots of sensitivity for these 4 curves
Gamma Variation - Unimportant Notes
Shown are also cloud representations of the data in question
Even the largest among the displacements which we apply is not obvious enough to catch the eye
Difference, effect of constant goes away
Values of A->A and A->B change, but not their difference
Gamma Variation and its Effect
Synthetic Data clouds in Space
Offset of 5 units in axes X, Y, Z for visualisation purposes
Larger Sample: Synthetic Data clouds in Space
Offset of 5 units in axes X, Y, Z for visualisation purposes
Sensitivity Curves and Gamma Variation
Data could be re-used for experiments that exploit large data sample
Fortunately, one can get a large (~85MB) data dump before quitting MATLAB
Able to re-use the data generated and produce sensitivity curves and the like rather quickly
Wrong Gamma Plots
This plot is wrong as it shows the wrong parameter altered
kNN and Gamma - Effects on Entropy Sensitivity
Realised that we had the wrong parameter changed in the sensitivity plots
Rather than changing Gamma, the program was changing K (as in kNN), which is still intersting
Finally produced correct sensitivity plots which demonstrate that there is no change in sensitivity, even if one goes as high as Gamma=10
Still have some nice plots to show the effect of varying K (plots I produced by mistake)
The functions take very many arguments, which makes them confusing to use (later to be corrected and improved)
The Variation of K - Low Values
K in the range 1-4
The Variation of K - Larger Values
K assigned the value of 2, 4, 5, 8, and 10
Early Observations
I had a hard time getting encouraging results
The curves that we get in practice differ from these which we saw in the simple example
Among the causes:
As indicated before, distances from the synthetic to the the training set are much larger than synthetic-synthetic distances
The value of Alpha is verging one because of the nigh number of dimensions
In the equation of entropy we multiply by 1/(1-Alpha), which is going to be a huge number
Distribution Size Increase and Resultant Entropy
Dealing with just 100 synthetic images
Thoughts About Results
Notes about the results refer to both the real and mockup case
Distribution expands -> inner distances increase, unlike positional movement of both samples
Synthetic to training likewise, reasons is the same as above
Training images are more 'different' from synthetic set to itself, thus the decline in the overall entropy (pace of change differs)
Overall Entropy
Entropy of Synthetic Distribution to Self
Entropy of Synthetic Distribution to Training Set
The Effect of Increasing One Distribution's Size
Specificity in the Corresponding Set of Experiments
More on Entropy Calculations
Obtained almost all the matrix data
Hoping that nobody switches off any of the machines at any point throughout the that time
We would be on the 'safe side' otherwise, in terms of data collection
Sent Carole all the paper files, programs, and data (about 100 MB)
Rechecked the experimental results that were sent beforehand and all seemed okay
Drawn a diagram to explain this and accompany previous correspondence, which was generated in haste
Reasoning About the Behaviour of Entropy Plots
Combined Entropy Curves
All the entropy curves with error bars, which are derived from the variation in graph length
Entropy Curves from the Real Data
Entropy and Specificity Combined Figure
Shown in this figure are the entropy and Specificity of the real data
No scaling of any kind has been applied
Entropy and Specificity (Real Data)
Plots and numbers are not scaled
Resources Available
Needed brute-force for calculation that are more statistically valid
Finally I got hold of over 30 computers, so I began running 3 more Si's for each value of displacement
We could end up with 4 Si's (148 images) for each sample point
Progress on Si's, Distribution Size Variation
Obtained a full second batch of Si's and verified that all matrices are okay
With our artificially-generated distribution data, I re-ran the experiments which involve distribution size. I happened to find that, as I suspected, a single apostrophe was the cause for the curves not meeting at 0. As you can see, the tails now meet. I vary the size of the dynamic distribution from %10 to 100% (identical size). By mirroring the image, you can see the graph that you otherwise expected to see. One distribution shrinks rather than grows
This is actually similar to what we saw in initial plots, which I produced from the real data 3-4 weeks ago (the two curves did not meet though)
Varying the Size of One Among 2 Distributions
Older Plots of Entropy
These are probably ill-computed
The Effect of Varying Gamma
Same as before, but a variety of assignments to Gamma are investigated.
3 Entropy 'Components' and Gamma Variations
Explanation of Results
The following might be worth pointing out:
There is a reason why, for the real data, the entropy will never start at 0
You are not comparing like with like
The distribution which is the training set and the distribution which is the synthetics derived from that optimal model are different
The model is lossy after all
Thus, the old plots for real data remain valid
They also lie in agreement (logically at least) with the mockup case
Entropy in Expanding Distribution
An interesting plot which shows the intersection of the two entropy values
Expanded Range for Distribution Size Variation
Deadlines and Milestones Ahead
TMI submission
ISBI Registration (and camera-ready version)
MICCAI deadline March 10th - suitable for introduction of entropy as a substitute for Specificity and Generalisation
MIUA deadline: April 24th
January 27th - a 15-minute Monday talk to make up for absence in Student Talks
Deadlines and Milestones Ahead - Ctd.
February 10th-14th: CVPR preliminary decisions and rebuttal period, finalised in March