Computer Vision News - November 2016

64 Computer Vision News Tool Results: Conclusion: In term of classification the two clusters start to emerge only from perplexity 50 and above. In terms of topology t-SNE doesn’t preserve the structure of those clusters (Cluster sizes in a t-SNE plot mean nothing, as also stated by van der Maaten!), t-SNE greatly exaggerates the size of the smaller group of points, the outer group becomes a circle; it would be easy to misread these outer points as a one-dimensional structure. And only at perplexity 100 thigs become clearer. Dataset-7: Dataset of three Gaussian clusters with 200 points each. Results: Conclusion: As usual, at perplexity 30 (within the range) the three clusters are observed and distances between clusters might not mean anything. The examples demonstrate there are many proven, inspiring results. However, parameters are very critical and if not properly tuned might lead to misleading results. Real-world data is likely to include multiple clusters with a varying number of elements. In this case, it may be no single perplexity value will capture distances across all clusters – too bad perplexity is a global parameter. This problem may prove an interesting area for future research. Sources: (1) Wattenberg, et al. “How to Use t-SNE Effectively”, Distill, 2016. (2) https://lvdmaaten.github.io/tsne/ (3) L.J.P. van der Maaten and G.E. Hinton. Visualizing High-Dimensional Data Using t-SNE. Journal of Machine Learning Research 9 (Nov): 2579-2605, 2008. Tool