Untrained neural networks in data-scarce smart manufacturing: a paradigm for data-efficient industrial intelligence

The progression towards Industry 4.0 has intensified the need for intelligent systems in manufacturing. However, the prevailing deep learning paradigm is often hindered by its reliance on large, labeled datasets, which are frequently unavailable in industrial settings where defect data is inherently rare and costly to acquire. This review introduces Untrained Neural Networks (UNNs) as a powerful, data-efficient alternative. UNNs leverage the intrinsic architectural biases of neural networks, particularly Convolutional Neural Networks, as an implicit prior to solve complex problems without any pretraining. This approach, which requires no training and is exemplified by the Deep Image Prior (DIP) framework, enables tasks like image reconstruction and feature extraction using only a single data sample. This paper reviews the core principles of UNNs, including their inherent characteristics and their integration with physics-informed models and training-free Neural Architecture Search. We then detail their primary applications in solving industrial inverse imaging problems, such as computational microscopy and non-destructive testing, as well as in performing unsupervised anomaly detection. Furthermore, we analyze the distinct advantages of UNNs for smart manufacturing through the lens of the Industrial Internet of Things (IIoT). The review concludes by identifying key challenges, including reliability and scalability, and proposing future research directions focused on enhancing trustworthiness, developing specialized architectures, and enabling adaptation to dynamic industrial environments. This work highlights the significant potential of UNNs to lower the barrier for AI adoption, offering a flexible and cost-effective solution for industrial scenarios where data is scarce.