The reusability prior: comparing deep learning models without training

Abstract Various choices can affect the performance of deep learning models. We conjecture that differences in number contexts for model components during training are critical. generalize this notion by defining reusability prior as follows: forced to function diverse not only due data, augmentation, and regularization choices, but also design itself. focus on aspect introduce a graph-based methodology estimate each learnable parameter. This allows comparison models without requiring any training. provide supporting evidence with experiments using cross-layer parameter sharing CIFAR-10, CIFAR-100, Imagenet-1K benchmarks. give examples share parameters outperforming baselines have at least 60% more parameters. The graph-analysis-based quantities we introduced align well results, including two important edge cases. conclude provides viable research direction analysis based very simple idea: counting