COMPARATIVE ANALYSIS OF DYNAMIC ADAPTATION OF NEURAL NETWORKS

Abstract

Purpose. The study aims to systematize methods for dynamic adaptation of neural networks and to determine how structural pruning, architecture growth, combined restructuring, and conditional execution affect classification accuracy, parameter count, and estimated forward-pass cost. Methods. The study distinguishes structural pruning, architecture growth, combined restructuring, and conditional execution. For each group, it describes the mechanisms of structural change, the approach to complexity control, and the treatment of computational constraints. Experimental comparison was performed on image datasets of different complexity with fixed training, validation, and test subsets. Classification quality was assessed by final and best validation accuracy and training dynamics; model compactness was assessed by parameter count and the proportion of nonzero weights; computational cost was assessed by the number of forward-pass operations. Results. Under short training protocols, structural pruning and moderate compression generally provided the most balanced trade-off among accuracy, model compactness, and computational cost. More aggressive compression reduced model size but could decrease accuracy. Conditional execution did not achieve competitive accuracy on the more complex image dataset, indicating that such methods are sensitive to the training budget and the configuration of the control mechanism. Conclusions. For the first time, four groups of dynamic neural-network adaptation methods were systematically compared within a unified short experimental protocol. The results show that method selection should be guided by task requirements, data complexity, the base architecture, and available resources. Practical model assessment should complement accuracy and compactness measures with measurements of actual inference latency and energy consumption on target hardware.