We'll take a look at what t-sne does with the codewords produced by vggnet on the xtcav images.
- Grabbed 4000 codeword2
- about 2000 lasing, 2000 no-lasing
- grabbed the t-sne python code from https://lvdmaaten.github.io/tsne/,
- another good reference for t-sne: http://colah.github.io/posts/2014-10-Visualizing-MNIST/
- first eliminated the dead neurons - there were 1023 dead neaurons (at least from those 4000 samples, should check all)
- run tsne with default parameters (perplexity=30.0)
We get this image:
Interesting to see how well separated they are, it is odd to see a couple yellow no-lasing samples show up in the blue lasing area, shots in the no-lasing run should never exhibit lasing features. However vice-versa, makes sense. There are shots in the lasing run that did not lase, so we would expect some blue mixed with the yellow. It is interesting that most all of the blue that is in the yellow cluster is together - it suggests that the lasing shots that failed to lase have some other structure that links them together, and keeps them on the outskirts of the general no-lasing clump. It would be interesting to compare this plot to a plot of t-sne applied to the raw images, before being turned into vgg16 codewords.
Things to investigate
- Different choices for perplexity
- More data
- t-sne with the enPeaksLabels - the 0,1,2,3
- Running t-sne from different points in vgg16 net, ie, try
- both codeword1 and codeword2, the 8192 final values
- just codeword1
- just the output of convolutions
- Run t-sne directly on the input images
- Look up the input images, what do the lasing images in the blue blob that lives with the no-lasing have in common? Recall the 'lasing' was filtered, we have to measure a e1 or e2 in the enPeaksLabel, are those just mislabelings? They have no lasing to notice? Or is there some other structure in them?