That's very reminiscent of the idea behind the SAM (Sharpness Aware Minimization) family of optimizers.