Researchers have introduced Shallow Universal Polynomial Networks (SUPNs), a novel architecture designed for efficient function approximation. Unlike deep neural networks (DNNs) and Kolmogorov-Arnold Networks (KANs) that often require significant depth and width, SUPNs maintain a single hidden layer. This design enhances the expressivity of each unit by incorporating a learnable multivariate polynomial within its activation function, thereby reducing parameter counts and training costs while preserving a shallow computational graph. The research establishes that SUPNs can achieve universal approximation by increasing width or polynomial complexity and demonstrates their ability to approach optimal approximations with explicit parameter derivations and convergence rates. Empirical evaluations across over 16,000 trained models show SUPNs frequently outperform DNNs and KANs in approximation error and variability at comparable parameter counts, supporting their utility for efficient function approximation when target structures are unknown. AI
IMPACT This new architecture could lead to more efficient AI models with lower training costs and reduced variability in generalization error.
RANK_REASON Academic paper introducing a new model architecture. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Deep Neural Networks
- Kolmogorov-Arnold Networks
- Shallow Universal Polynomial Networks
- Zachary Morrow
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
