<ul><li>A new training framework has been introduced for training Polynomial Neural Networks (PNNs) effectively, crucial for privacy-preserving inference via Homomorphic Encryption.</li><li>The framework addresses challenges such as limited model expressivity with low-degree polynomials and numerical instability with high-degree polynomials.</li><li>Key innovations in the framework include a Boundary Loss that penalizes activation inputs outside a stable range and Selective Gradient Clipping to manage gradient magnitudes.</li><li>The framework enables training of PNNs with polynomial degrees up to 22, showcasing stable training, strong performance, and compatibility with Homomorphic Encryption across various datasets.</li></ul>

A Training Framework for Optimal and Stable Training of Polynomial Neural Networks

Discover more