June 15th, 2006 ___________________________________________________________ MIUA 2006 * Edited to paper, addressing the reviewers' comments in the process * Passed it around for feedback * The reviewers' comments are included in the previous weekly report * These are certainly worth discussing in depth * Some MIUA reviewers raised points that were similar to the ones pointed out by TMI reviewers ___________________________________________________________ Training Needs Meeting, Discussion with Advisor * Plans for thesis submission and what is yet to come * Accommodation is a concern as my contract runs out in August, can be renewed (room reserved until tomorrow) ___________________________________________________________ Experimental Work * 7 new instantiations were being generated by 7 computers * Came to realise that MATLAB's built-in seeding is inexistent * Started sessions on several independent computers and the first random number remained identical across all * Vlad - discussion about a fitting algorithm which accounts for the error bars and gives insight into measure sensitivity * The method is described in "Numerical Recipes in C". ___________________________________________________________ Items Worth Discussing * Updated/modified fine-level details in TMI submission; more notes on the printed copy; need to re-run 10 instantiations * In our last meeting I wanted to discuss whether it would be sensible for me to write my thesis while (re-)running these experiments * I also wanted to discuss the future as my accommodation contract will have come to an end by August * I sent a correction to a TMI figure. I used grey shades to separate the regions vertically. The labels otherwise 'collide', which would confuse the observer. ___________________________________________________________ Gentle Request for Tanimoto Overlap (Bill) * Random seeding in MATLAB meant that while we used 10 instantiations in our validation experiments, there was repetition, so we only have 3 truly separate instantiations * We only came to realise this more recently and, while the results in general are not expected to change, we want to do the experiments properly * Therefore, we are generating another 7 instantiations * Bill is computing the Tanimoto overlap for these, in the same way as before * Long sequences of "random" numbers needed * I was referred to the Mersenne Twister sequence: http://www.math.sci.hiroshima-u.ac.jp/~m-mat/MT/emt.html * Resolved the issue * Generated and put up 7 new instantiations, which are formatted in the same way as before * Converted all the data to Analyze (4 GB) * Compressed everything, which shrank it by a factor of 10 * Results delivered a few weeks later * We need to justify the validation results for ourselves, as well as modify the figures. ___________________________________________________________ Recent Work * Primarily work on the thesis, chapter by chapter; I have all the graphics in place, as well as very large portions of text * Producing all the file (videos) for the accompanying CD-ROM * Bill has finished evaluating the new instantiations * Model-based Evaluation is carries out over here, progressing faster than I imagined it would * BMVC - they have had delays in making any decisions (already 3 days overdue) * TMI decision is due soon (probably by the end of month); we will have the new instantiations by then * Bill's paper on overlap-based assessment was accepted to TMI * MIUA presentation - I am assuming it can be pretty much the same as in MIAS-IRC 2005 (same work is described) * A group from Lunds contacted me yesterday, showing interest in our work on ground-truth-free NRR assessment ___________________________________________________________ Experiments - Progress * Repeat experiments - random instantiations * Looked at some preliminary results * The numbers I have taken a glimpse at suggest that they will be no unpleasant surprises * The rise in values seems monotonic, at least for instantiations that I looked at ___________________________________________________________ Sensitivity Calculations * You: "The idea of fitting a straight line through the data makes sense, though Tim's suggestion that we should measure the ratio between the slope and the error in the slope does not make sense (he agrees), since any sensible definition of sensitivity needs dimensions (Warp Magnitude)^-1" * I spoke to Vlad on several occasions and I believe he implemented a sensitivity estimator * He is now just waiting for the numbers * I have about 2.5 instantiations left (about a week to complete) to process and Bill has just sent his results