Computer Vision News - November 2016

Dataset-4: 500 points drawn from a unit Gaussian distribution in 100 dimensions. Results: Conclusion: Perplexity 2 seems to show dramatic clusters, however, at perplexity 30 (inside the range 5-50) points seem evenly distributed. These phenomena are not bad, actually they are telling us useful things about high-dimensional normal distributions, which are very close to uniform distributions on a sphere. Dataset-5: Two clusters of 75 points each in 2D, arranged in parallel lines with a bit of noise. Results: Conclusion: Again, at perplexity 30 (inside the suggested range) the two clusters are observed. However, there’s a subtle distortion: the lines are slightly curved outwards in the t-SNE diagram. The reason is that, t-SNE tends to expand denser regions of data – the middles of the clusters have less empty space around them than the ends, thus the algorithm magnifies them. Dataset-6: Two groups of 75 points in 50 dimensional space which are sampled from symmetric Gaussian distributions centered at the origin, but one is 50 times more tightly dispersed than the other. The “small” distribution is in effect contained in the large one. Computer Vision News Tool 63 Tool