As others have pointed out, SGD requires differentiability. The next best match I can think of is that you're actually thinking of subgradient methods, which mo

As others have pointed out, SGD requires differentiability. The next best match I can think of is that you're actually thinking of subgradient methods, which mostly see use in convex optimization problems.