How Neural Networks Extrapolate: From Feedforward to Graph Neural Networks.

We study how neural networks trained by gradient descent extrapolate, i.e., what they learn outside the support of training distribution. Previous works report mixed empirical results when extrapolating with neural networks: while multilayer perceptrons (MLPs) do not extrapolate well in simple tasks, Graph Neural Networks (GNNs), a structured network with MLP modules, have some success in more complex tasks. We provide a theoretical explanation and identify conditions under which MLPs and GNNs extrapolate well. We start by showing ReLU MLPs trained by gradient descent converge quickly to linear functions along any direction from the origin, which suggests ReLU MLPs cannot extrapolate well in most non-linear tasks. On the other hand, ReLU MLPs can provably converge to a linear target function when the training distribution is "diverse" enough. These observations lead to a hypothesis: GNNs can extrapolate well in dynamic programming (DP) tasks if we encode appropriate non-linearity in the architecture and input representation. We provide theoretical and empirical support for the hypothesis. Our theory explains previous extrapolation success and suggest their limitations: successful extrapolation relies on incorporating task-specific non-linearity, which often requires domain knowledge or extensive model search.