No Train, all Gain: Self-Supervised Gradients Improve Deep Frozen Representations | Read Paper on Bytez