← All questions
Retrieval & EmbeddingsMedium· Sparsity· Regularization· Optimization

Why can't you simply minimize the number of nonzeros in...

Minimizing the L0 norm directly is impossible because it is piecewise constant. L1 serves as a convex relaxation of the count, while FLOPS acts as a relaxation of the retrieval cost using mean activation as a proxy for activation probability. It is important to note that the ReLU activation function is what actually supplies the exact zeros in the output vector.

Code playground (Python)