This paper introduces Quad-Directional Modular Networks (QDMN), a novel neural architecture that treats "Up" and "Down" directions as first-class processing routes alongside traditional forward and backward passes. QDMN operates as an explicit multi-route processing fabric, allowing for complex information flow between modular units. The research explores whether expanding the directional dimensionality of neural networks can improve representational capacity and stability. Key highlights include: Architectural Design: Implementation of a contractive refiner to ensure stability across deep multi-directional sweeps. Multi-Route Processing: Analysis of how information propagates through the network when given explicit directional choices. Experimental Results: Comparative studies against standard MLP baselines across varying depths, demonstrating the trade-offs between directional complexity and model performance. This work is presented as an exploratory draft to investigate non-traditional processing directions in neural network design.
Sohan Poudel (2026) studied this question.