Deep learning applications and interpretation in signal processing and magnetic resonance

Abstract

With the advent of Artificial Intelligence, Machine Learning and growing computer power, applications of these are finding widespread use within multiple scientific and data driven applications [1]. Neural networks are a sub-division of Machine Learning, known as Deep Learning [2], both the former and latter falling under the umbrella of the term Artificial Intelligence [3]. Neural networks have been used in a multitude of problems, including image [4], speech and voice recognition [5], time series forecasting [6] and signal processing [7]. They have now found their way into solving complex problems within molecular biology, namely in DEER or Double Electron-Electron Resonance, also known as PELDOR (Pulsed Electron Double Resonance) [8]. The networks are seemingly able to invert the Fredholm integral equation which represents the distance distribution between two isolated spin electron pairs. DEER data has been previously processed using Fourier transform methods [9-11], analytical solutions [12], Tikhonov regularisation [13, 14] and other regularisation methods [15, 16]. The challenge now is to figure out exactly how these networks perform such tasks. The importance of this reaches beyond the scope of physical and biological chemistry, it is a universal criticism of using deep learning models; they are generally known as “black boxes” [17]. We know what they are doing on a surface level, in terms of vector and matrix multiplication along with the calculus involved in gradient descent and backpropagation in their training, but exactly what physically motivated mathematical transformations they perform often remains a mystery [17-19].

More information

Identifiers

ORCID for Ilya Kuprov: orcid.org/0000-0003-0430-2682

Catalogue record

Export record