Extended Data Table 1 Results of ablation study and methods’ comparison. Ablated architectures are obtained by removing the learnable local receptive field (LLRF) and/or the gated self-attention (GSA) from MAGIK. Reported metrics correspond to averages over 20 runs for linking and over 5 runs for the other cases. Best-in-class performances are reported in bold. The dashed lines correspond to the metric values obtained for the baseline model

From: Geometric deep learning reveals the spatiotemporal features of microscopic motion