No Train, all Gain: Self-Supervised Gradients Improve Deep Frozen Representations Walter Simoncini 1, Andrei Bursuc