Beyond Linearization: On Quadratic and Higher-Order Approximation of Wide Neural Networks
This addresses the insufficiency of NTK theory for explaining neural network performance, offering a novel theoretical framework for understanding and improving training in over-parametrized networks.
The paper tackles the gap between neural network performance and linearized models by introducing randomization to couple over-parametrized networks with quadratic and higher-order approximations, proving improved sample complexity bounds that can exceed NTK results by a dimension factor under mild assumptions.
Recent theoretical work has established connections between over-parametrized neural networks and linearized models governed by he Neural Tangent Kernels (NTKs). NTK theory leads to concrete convergence and generalization results, yet the empirical performance of neural networks are observed to exceed their linearized models, suggesting insufficiency of this theory. Towards closing this gap, we investigate the training of over-parametrized neural networks that are beyond the NTK regime yet still governed by the Taylor expansion of the network. We bring forward the idea of \emph{randomizing} the neural networks, which allows them to escape their NTK and couple with quadratic models. We show that the optimization landscape of randomized two-layer networks are nice and amenable to escaping-saddle algorithms. We prove concrete generalization and expressivity results on these randomized networks, which lead to sample complexity bounds (of learning certain simple functions) that match the NTK and can in addition be better by a dimension factor when mild distributional assumptions are present. We demonstrate that our randomization technique can be generalized systematically beyond the quadratic case, by using it to find networks that are coupled with higher-order terms in their Taylor series.