Extended Data Fig. 4: Ablations on different network components of the GNN.

(a-d) Effect of update pruning due to max pooling. We interpret max pooling as a kind of event filter. In (a-b) we show an example of aggregated events before (a) and after (b) filtering. This filter acts as a saliency detector, only letting through events with “new information”, and removing redundant events in high event rate regions. This results in a more uniform distribution of events (c). We can control the filter strength by modulating the number of output features, c. As seen in (d), increasing c increases both computation and mAP. However, mAP growth drastically reduces in slope after c = 24. The dot size is proportional to c, and ϕ measures the proportion of updates that pass through the filter. In our baseline setting with c = 16, we see that only 27% of updates pass the first max pooling layer. (e) Features affecting computational complexity. (f) Features affecting accuracy.