A Publication of the Signals and Systems Research Group receives a Hojjat Adeli Award

A Publication of the Signals and Systems Research Group receives a Hojjat Adeli Award HU
The International Journal of Neural Systems is a monthly, transdisciplinary, peer-reviewed, international scientific journal (D1, impact factor: 8.0), focusing on the information processing of natural and artificial neural systems, and touching on the subjects of machine learning, computational neuroscience, and neurology.

In honour of the journal’s editor-in-chief, the World Scientific Publishing Company established the Hojjat Adeli Award in 2010, which is awarded annually to the authors of the most innovative articles published in the journal. In 2023, the prize was awarded to the Signals and Systems Research Group hosted by the Department of Numerical Analysis at the ELTE Faculty of Informatics for their publication:

P. Kovács, G. Bognár, C. Huber, and M. Huemer, VPNET: Variable Projection Networks, International Journal of Neural Systems, 32:1, 2150054 (19 pages).

The paper was written as a result of joint research carried out at the ELTE Faculty of Informatics, the Institute of Signal Processing at the Johannes Kepler University Linz, and the department of Embedded AI at the Silicon Austria Labs research centre, and was published in the form of an open-access article. To facilitate the reproducibility of the results, we have published the data of the article, as well as the NumPy and PyTorch implementations.

The award-winning publication presents a novel model-driven neural network architecture developed by integrating parameterized orthogonal transformations. Model-driven neural network design is a popular subject of the so-called Explainable AI (XAI) investigations combining a mathematical model-based approach with data-driven algorithms of artificial intelligence. In the publication above, numerical methods for nonlinear least-squares problems – often discussed in physics and engineering science – have been embedded in fully connected neural networks. The variable projecting network (VPNet) obtained in this way is a model-driven neural network, which comprises the numerical and mathematical heuristics needed to solve the target task. This offers several benefits, such as increased performance, improved estimation accuracy, and the explainability of the decisions made by the model.

Since the publication of the basic concept, we have tested the effectiveness of VPNnets in several applications related to interdisciplinary informatics, such as the classification of ECG signals by arrhythmias, the EEG-based discrimination of visually evoked potentials (VEP), and the detection of potholes via rubber sensor signals. Several MSc students have been involved in the study of the potential practical applications of the network in the form of master’s theses and papers submitted to the Hungarian Students’ Scholarly Circle (TDK) contest through the Model-Building R&D lab run by the Department of Numerical Analysis at ELTE. Interested students are welcome to join this work this year as well.