Fig. 3: Empirical data on diversity of within-benchmark dynamics.
From: Mapping global dynamics of benchmark creation and saturation in artificial intelligence
a Diversity of within-benchmark dynamics patterns observed for the metrics across all benchmarks from NLP and computer vision. For assignments of individual benchmarks among the clusters see Supplementary Data 5). b Top-10 most similar trajectories to predefined gold trajectories representing linear growth (\(f(x)=x\) where \(\left\{x{\mathbb{\in }}{\mathbb{N}}1\le x\le 50\right\}\)), early saturation (\(f(x)=-1/x\) where \(\left\{x{\mathbb{\in }}{\mathbb{N}}1\le x\le 50\right\}\)) and stagnation followed by breakthrough (\(f(x)=-1/x\) where \(\left\{x{\mathbb{\in }}{\mathbb{N}}{\mathbb{-}}50\le x\le -1\right\}\)). As similarity metric we use the euclidean distance between trajectory and gold function.
