On the effects of similarity metrics in decentralized deep learning under distributional shift