Visualizing High Dimensional MNIST Dataset with PCA and t-SNE

In this notebook I am trying to visualize natural clusters of digits in MNIST dataset.
MNIST dataset has 784 dimensions so to visualize it i am converting it in to 2 dimensions using two most popular dimensionality reduction technique called PCA and t-SNE.

this is the link to the notebook:

1 Like

I would suggest implementing an autoencoder with smaller than input latent space vector and use this as a base for analysis.
You could also try different latent space sizes and see how it affects the outcome.

It’s still interesting that analysing the images effectively pixel by pixel (784 dimensions) gives such nice results (especially t-SNE). Although, with AE, I suspect the PCA/t-SNE algorithms will run faster.

1 Like

Actually I am newbie in this field so don’t know much about AE but i will definitely learn and try it on this dataset thanks. : )