Jakub Vrábel

J. Vrábel, E. Képeš, P. Nedělník et al.

The mutual incompatibility of distinct spectroscopic systems is among the most limiting factors in Laser-Induced Breakdown Spectroscopy (LIBS). The cost related to setting up a new LIBS system is increased, as its extensive calibration is required. Solving the problem would enable inter-laboratory reference measurements and shared spectral libraries, which are fundamental for other spectroscopic techniques. In this work, we study a simplified version of this challenge where LIBS systems differ only in used spectrometers and collection optics but share all other parts of the apparatus, and collect spectra simultaneously from the same plasma plume. Extensive datasets measured as hyperspectral images of heterogeneous specimens are used to train machine learning models that can transfer spectra between systems. The transfer is realized by a pipeline that consists of a variational autoencoder (VAE) and a fully-connected artificial neural network (ANN). In the first step, we obtain a latent representation of the spectra which were measured on the Primary system (by using the VAE). In the second step, we map spectra from the Secondary system to corresponding locations in the latent space (by the ANN). Finally, Secondary system spectra are reconstructed from the latent space to the space of the Primary system. The transfer is evaluated by several figures of merit (Euclidean and cosine distances, both spatially resolved; k-means clustering of transferred spectra). The methodology is compared to several baseline approaches.

Jakub Vrabel, Ori Shem-Ur, Yaron Oz et al.

We extend the concept of loss landscape mode connectivity to the input space of deep neural networks. Mode connectivity was originally studied within parameter space, where it describes the existence of low-loss paths between different solutions (loss minimizers) obtained through gradient descent. We present theoretical and empirical evidence of its presence in the input space of deep networks, thereby highlighting the broader nature of the phenomenon. We observe that different input images with similar predictions are generally connected, and for trained models, the path tends to be simple, with only a small deviation from being a linear path. Our methodology utilizes real, interpolated, and synthetic inputs created using the input optimization technique for feature visualization. We conjecture that input space mode connectivity in high-dimensional spaces is a geometric effect that takes place even in untrained models and can be explained through percolation theory. We exploit mode connectivity to obtain new insights about adversarial examples and demonstrate its potential for adversarial detection. Additionally, we discuss applications for the interpretability of deep networks.