Sparse coding
Representing data with few active coefficients in a learned dictionary; an unsupervised model related to neural feature learning and compression.
Sparse coding learns a dictionary of features so each input is reconstructed from only a few active coefficients. The representation is sparse: most coefficients stay near zero, which can yield interpretable parts and efficient codes.
The ICML 2009 GPU paper by Raina, Madhavan, and Andrew Ng treated sparse coding alongside deep belief networks as a workload that benefits from parallel GPU hardware (article). The idea connects classical unsupervised learning to later deep learning feature learning.