This problem was studied by the National Research Council, Informatics Group. With three different sets of data, they obtained about 90, 80, 70% classification accuracy with their 'proprietary' algorithm. We have their reports with methods and results in detail. We disclose them here for your information:The ReportAddendum
1. I presume that most of the analyses performed on our data thus far was focused on the frequency domain. It should be noted that the analyses with high accuracy rates by the NRC Informatics Group specifically focused on the time domain
, i.e. signature features of a given substance as a function of time of data acquisition after exposure of the sensor to the material of interest.
2. In the most recent analysis of our data by the NRC, we were foretold by the Group that they would try to improve their accuracy by making classifications on the basis of frequencies. We had predicted that the algorithms chosen, like wavelets, were inappropriate. Indeed, their accuracy was in the 0.3 range, i.e. equivalent to random noise and comparable to the values in the Leaderboard.
3. Have you looked at the Sigma Plots
of the contest data?
There are many consistent features across different rows of panels: size of 'peaks', degree of uniformity of the various sizes of peaks, the regularity and distribution of unusually large peaks for a given row of images, the overall coherence of the pattern across each image of a row of images and across the entire five panels in the row; and others. That the patterns, for a given row (substance) are consistent across the five panels (representing signatures across the whole period of data acquisition), implies a coherence in signal signatures. Consequently, one should note the consistency/coherence across each row of images. I would expect that image analysis procedures would dissect these types of differences, and others, to provide a digital equivalent of the images.
Let me point out that imposition of coherence by exposure of the sensor to a substance is seen in every instance and immediately. This was noted by us with the original analog output of the sensor, i.e. a paper chart recorder, and is characteristic of "stochastic resonance"
, which appears to be the basic mechanism underlying our observations: a very weak signal can only be detected by its modulation of the background noise
. Indeed, a characteristic of this phenomenon: a weak signal imposes coherence on the background noise pattern, and as a consequence, noise reduction