Session Ready
Exercise

Perplexity of MNIST sample

The perplexity parameter indicates the balance between the local and global aspect of the input data. The parameter is an estimate of the number of close neighbors of each original point. Typical values of this parameter fall in the range of 5 to 50.

We will generate three different t-SNE executions with the same number of iterations and perplexity values of 5, 20, and 50 and observe the differences in the K-L divergence costs. The optimal number of iterations we found in the last exercise (1200) will be used here.

The mnist_sample dataset and the Rtsne package have been loaded for you.

Instructions 1/3
undefined XP
  • 1

    Fix the seed to 1234. Generate tsne_output using a perplexity of 5 and 1200 iterations. Show the itercosts values.

    • 2

      Fix the seed to 1234. Generate tsne_output using a perplexity of 20 and 1200 iterations. Show the itercosts values.

    • 3

      Fix the seed to 1234. Generate tsne_output using a perplexity of 50 and 1200 iterations. Show the itercosts values.