Abstract
Deep neural networks (DNNs) have achieved remarkable success across diverse machine learning tasks through complex pattern recognition and continuous optimization. Despite these advances, DNNs often require extensive manual tuning of architectural parameters, such as layer depth and neuron counts, which are highly problem dependent. This lack of automation not only increases development time and computational cost but also limits the scalability and generalization of neural architectures in dynamic and data-intensive environments. To address these challenges, this thesis introduces a novel automated architecture design framework based on Structurally Adaptive Neural Networks (SANets). The proposed approach begins with training a fully connected DNN (FC-DNN) initialized using conventional heuristics for hidden layer sizes. SANet then performs a two-phase structural adaptation process. In the first phase, a lazy magnitude-based pruning strategy eliminates weak connections and isolated neurons, producing a series of partially connected DNN (PC-DNN) models. In the second phase, a fine-tuning procedure refines the architecture and converges toward an efficient configuration. This adaptability bridges the gap between static model design and environments that demand continuous learning and optimization. Consequently, SANets are particularly well-suited for domains characterized by uncertainty, temporal variation, and high-dimensional data dependencies. Applied to marine datasets, SANet demonstrates superior efficiency, scalability, and predictive accuracy compared to conventionally pruned FC-DNNs. Overall, the proposed framework advances the automation and adaptability of DNN design, providing a robust and practical foundation for real-world machine learning applications operating in dynamic environments.