Training a graph neural network on multiple objectives usually means blending conflicting gradients at every step. Instead of compromising among parameter updates from different training objectives, ControlG allocates capacity to objectives sequentially and dynamically via PID



