Differentiable relaxation makes it sound like Rowhammer backpropagation, but I'm fairly certain that's not what they're talking about.