Computer Vision News - November 2016

The location of the points y i are set by minimizing the Kullback–Leibler divergence of the distribution Q from the distribution P, that is: Note, most often t-SNE starts with a preprocessing step using PCA. This speeds up the computation of pairwise distances between the data-points and suppresses some noise without severely distorting the interposing distances. Now let’s start with the fun part, analyzing several examples. We will start with two examples given by Laurens van der Maaten in the original t-SNE paper. Follow by a set of examples from Wattenberg, et al. “How to Use t-SNE Effectively”, Distill, 2016. Dataset-1: The well-known MNIST data set contains 60,000 handwritten images. t-SNE Results: Conclusion: t-SNE constructs a map in which the separation between the digit classes is almost perfect. In addition, a detailed inspection of the t-SNE map reveals that much of the local structure of the data (such as the orientation of the ones) is well captured. Dataset-2: COIL-20 is a data set of 1440 toy images, each of size 128 × 128, in 20 classes. Computer Vision News Tool 61 Tool ( || ) = ≠ (