Abstract
A Transformer-based deep direct sampling method is proposed for solving aclass of boundary value inverse problem. A real-time reconstruction is achievedby evaluating the learned inverse operator between carefully designed data andthe reconstructed images. An effort is made to give a case study for afundamental and critical question: whether and how one can benefit from thetheoretical structure of a mathematical problem to develop task-oriented andstructure-conforming deep neural network? Inspired by direct sampling methodsfor inverse problems, the 1D boundary data are preprocessed by a partialdifferential equation-based feature map to yield 2D harmonic extensions indifferent frequency input channels. Then, by introducing learnable non-localkernel, the approximation of direct sampling is recast to a modified attentionmechanism. The proposed method is then applied to electrical impedancetomography, a well-known severely ill-posed nonlinear inverse problem. The newmethod achieves superior accuracy over its predecessors and contemporaryoperator learners, as well as shows robustness with respect to noise. Thisresearch shall strengthen the insights that the attention mechanism, despitebeing invented for natural language processing tasks, offers great flexibilityto be modified in conformity with the a priori mathematical knowledge, whichultimately leads to the design of more physics-compatible neural architectures.