The late 1980s through early 1990s saw gradient-based backpropagation emerge as the practical engine for training multilayer networks, unifying feature learning, representation construction, and function approximation across feedforward and recurrent forms. Researchers emphasized training signals that steer learning across layers, enabling deeper architectures to extract hierarchical structure while recurrent variants introduced online, time-aware updates that operate while networks run. Architectural strategies focusing on growth, tiling-like expansion, and compact representations accompanied domain-informed templates for pattern recognition and image understanding. Historical Significance: This period solidified backpropagation as the core optimization paradigm for neural networks, established the theoretical and practical viability of universal function approximation with multilayer structures, and laid groundwork for temporal learning with backpropagation through time and modular, mixture-inspired approaches that influenced later deep architectures.