Deep learning based ultrasonic reconstruction of rough surface morphology

This paper introduces a methodology to recover the morphology of a complex rough surface from ultrasonic pulse echo measurements with an array of equidistant sensors using the one dimensional convolution neural network (1DCNN). The neural network is trained by the datasets simulated from high-fidelity finite element simulations for surfaces with a range of roughness parameters and is tested on both numerical and real experimental data. To assess the performance of our proposed method, the rough surface reconstruction results from the deep learning approach are compared with those obtained from conventional ultrasonic array imaging methods. Unlike array imaging-based methods that require a large number of sensors (e.g., 128, 64 or 32), the deep learning-based method uses pulse echo signals and can achieve accurate results with much fewer sensors. The developed deep learning approach has the potential to enable low-cost, accurate, and real-time reconstruction of complex surface profiles.

Quantifying the separability of data classes in neural networks

We introduce the Generalized Discrimination Value (GDV) that measures, in a non-invasive manner, how well different data classes separate in each given layer of an artificial neural network. It turns out that, at the end of the training period, the GDV in each given layer L attains a highly reproducible value, irrespective of the initialization of the network's connection weights. In the case of multi-layer perceptrons trained with error backpropagation, we find that classification of highly complex data sets requires a temporal reduction of class separability, marked by a characteristic 'energy barrier' in the initial part of the GDV(L) curve. Even more surprisingly, for a given data set, the GDV(L) is running through a fixed 'master curve', independently from the total number of network layers. Finally, due to its invariance with respect to dimensionality, the GDV may serve as a useful tool to compare the internal representational dynamics of artificial neural networks with different architectures for neural architecture search or network compression; or even with brain activity in order to decide between different candidate models of brain function.