Engineering Applications of Neural Networks

Lazaros Iliadis, Ilias Maglogiannis, Serafí­n Alonso, Chrisina Jayne, Elias PimenidisCommunications 2023 · DOI · OpenAlex

Introduction

The 24th International Conference on Engineering Applications of Neural Networks (EANN 2023), held in León, Spain, serves as a critical barometer for the state of applied Artificial Intelligence. The proceedings, comprising 49 peer-reviewed papers, highlight a significant transition in the field: a move away from opaque "black-box" deep learning models toward architectures that are interpretable, ethical, and tailored for specific engineering domains. This shift is driven by the "deployment gap"—the challenge of implementing theoretical models in high-stakes, resource-constrained, and adversarial real-world environments.

Addressing the Deployment Gap

Engineering applications of AI face unique hurdles that distinguish them from general-purpose machine learning. Transparency is paramount; in sectors like medicine or aviation, "black-box" models lack the necessary trust for critical decision-making. Furthermore, engineering often involves edge devices (IoT) with limited computational power, requiring resource-efficient models. Finally, robustness is essential as AI integrates into industrial control and cybersecurity, where models must withstand adversarial attacks and generalize to unseen data patterns in real-time.

Methodological Innovations

To bridge this gap, the EANN 2023 proceedings showcase a variety of advanced strategies:

  • Interpretable-by-Design Models: Rather than relying on post-hoc explanations, new architectures utilize "prototypes" and kernel-like functions to ensure inherent transparency.
  • Federated and Transfer Learning: These techniques enable training on decentralized, privacy-sensitive data and allow for the adaptation of large pre-trained models to niche engineering tasks, such as marine debris detection.
  • Hybrid Architectures: Researchers are increasingly combining neural networks with classical engineering methods, such as fuzzy logic for medical diagnostics or ARIMA models for structural health monitoring.
  • Evolutionary Neural Architecture Search (NAS): Automated search algorithms are being used to optimize model structures for specific hardware constraints, balancing performance with energy efficiency.

Domain-Specific Applications and Datasets

The shift toward applied AI is reflected in the specialized datasets utilized across the conference papers. These move beyond standard benchmarks to address real-world engineering challenges:

  • Infrastructure: Analysis of cooling tower logs from CERN, seismic vulnerability assessments, and dam conductivity monitoring.
  • Healthcare: fMRI-based ADHD discrimination and signal processing for voice prostheses.
  • Environment and Energy: Remote visual inspection of offshore assets and multi-location wind energy forecasting.
  • Niche Engineering: Applications ranging from urban firefighting knowledge graphs to pothole detection and table tennis performance analysis.

Key Insights and Future Directions

A recurring theme, emphasized by invited speakers such as Plamen Angelov, is the necessity for "interpretable-by-design" AI in high-stakes engineering. Furthermore, the dual nature of AI in cybersecurity was highlighted: while it enhances malware detection, it also empowers attackers with sophisticated tools.

The conference also underscored the importance of "Hardware-Aware AI," focusing on techniques like network pruning to optimize energy efficiency for IoT environments. From an ethical standpoint, the 3rd Workshop on AI and Ethics (AIETH 2023) called for preemptive scientific action to mitigate the potential negative societal impacts of the 4th Industrial Revolution.

Conclusion

EANN 2023 demonstrated the successful implementation of neural networks across diverse fields, achieving notable results such as a 24% efficiency increase in industrial flow meters and improved zero-shot learning for autonomous driving. The conference maintained high standards with a 39% acceptance rate, featuring research from 19 different countries. However, challenges remain, particularly regarding the "robustness gap" and the scarcity of high-quality engineering data. Future research is expected to focus on semi-supervised and contrastive learning to maximize limited datasets, alongside the development of robust ethical and regulatory frameworks to keep pace with rapid technical advancements.

Summary updated

BibTeX
@inproceedings{iliadis2023engineering,
  author = {Lazaros Iliadis and Ilias Maglogiannis and Serafí­n Alonso and Chrisina Jayne and Elias Pimenidis},
  title = {{Engineering Applications of Neural Networks}},
  booktitle = {Communications},
  year = {2023},
  doi = {10.1007/978-3-031-34204-2},
}
Figure