Comment by zygy

Naive question: what's the intuition for how this is different from increasing the number of learnable parameters on a regular MLP?