<ul><li>Oversmoothing is a major limitation of Graph Neural Networks (GNNs) where input node features converge to a non-informative representation with bounded weights.</li><li>Backward oversmoothing is analyzed in this paper, indicating that backpropagated errors leading to gradient computations are also subjected to oversmoothing from output to input.</li><li>The interaction between forward and backward smoothing, especially with nonlinear activation functions, plays a key role in this phenomenon, leading GNNs to have many spurious stationary points.</li><li>This paper sheds light on the optimization landscape specific to deep GNNs and highlights the distinctive nature of this issue compared to other network architectures like Multi-Layer Perceptrons.</li></ul>

Backward Oversmoothing: why is it hard to train deep Graph Neural Networks?

Discover more