represent the data by K vectors.
This can also be viewed as the minimization of a loss function
can be thought of a reconstruction error.
in assignment, we minimize the assignments to minimize error (fix c, optimize i)
in other step, fix indexing of all datapoints, then update its center.
type of autoencoder.
latent representation = one hot vector
works on MNIST - kmeans learns numbers
can run on 8x8 image patches

Last Reviewed: 10/25/2025