A novel [[EQUATION]] approach to shape optimisation with Lipschitz domains

This article introduces a novel method for the implementation of shape optimisation with Lipschitz domains. We propose to use the shape derivative to determine deformation fields which represent steepest descent directions of the shape functional in the $W^{1,\infty}-$ topology. The idea of our approach is demonstrated for shape optimisation of $n$-dimensional star-shaped domains, which we represent as functions defined on the unit $(n-1)$-sphere. In this setting we provide the specific form of the shape derivative and prove the existence of solutions to the underlying shape optimisation problem. Moreover, we show the existence of a direction of steepest descent in the $W^{1,\infty}-$ topology. We also note that shape optimisation in this context is closely related to the $\infty-$Laplacian, and to optimal transport, where we highlight the latter in the numerics section. We present several numerical experiments illustrating that our approach seems to be superior over existing Hilbert space methods, in particular in developing optimal shapes with corners.

Application of supervised descent method for 2D magnetotelluric data inversion

The supervised descent method (SDM) is applied to 2D magnetotellurics (MT) data inversion. SDM contains offline training and online prediction. The training set is composed of the models generated according to prior knowledge and the data simulated by MT forward modeling. In the training process, a set of descent directions from an initial model to the training models is learned. In the prediction, model reconstruction is achieved by optimizing an online regularized objective function with a restart scheme, where the learned descent directions and the computed data residual are involved. SDM inversion has the advantages of (1) being more efficient than traditional gradient-descent methods because the computation of local derivatives of the objective function is avoided, (2) incorporating prior uncertain knowledge easier than deterministic inversion approach by generating training models flexibly, and (3) having high generalization ability because the physical modeling can guide the online model reconstruction. Furthermore, a way of designing general training set is introduced, which can be used for training when the prior knowledge is weak. The efficiency and accuracy of this method are validated by two numerical examples. The results indicate that the reconstructed models are consistent with prior information, and the simulated responses agree well with the data. This method also shows good potential to improve the accuracy and efficiency in field MT data inversion.

Visualizing Deep Networks by Optimizing with Integrated Gradients

Understanding and interpreting the decisions made by deep learning models is valuable in many domains. In computer vision, computing heatmaps from a deep network is a popular approach for visualizing and understanding deep networks. However, heatmaps that do not correlate with the network may mislead human, hence the performance of heatmaps in providing a faithful explanation to the underlying deep network is crucial. In this paper, we propose I-GOS, which optimizes for a heatmap so that the classification scores on the masked image would maximally decrease. The main novelty of the approach is to compute descent directions based on the integrated gradients instead of the normal gradient, which avoids local optima and speeds up convergence. Compared with previous approaches, our method can flexibly compute heatmaps at any resolution for different user needs. Extensive experiments on several benchmark datasets show that the heatmaps produced by our approach are more correlated with the decision of the underlying deep network, in comparison with other state-of-the-art approaches.