Comparative Analysis of Modular, Sequential, and Parallel Neural Architectures on Inference Accuracy and Latency

Abstract

This paper investigates three neural network paradigms: (1) sequential (hierarchical layer stacking), (2) parallel (dual-branch processing), and (3) modular (decomposition via neural chains). How do these architectural paradigms compare in final regression accuracy (MSE) and training time across tasks of increasing nonlinear complexity, when all other training variables are held constant?...