Learned Sparsity

A training technique where the model learns which dimensions of a vector are most important, automatically zeroing out less relevant ones.