SharpNet: Enhancing MLPs to Represent Functions with Controlled Non-differentiability
Researchers have developed SharpNet, a novel Multi-layer Perceptron (MLP) architecture designed to accurately represent functions with sharp, non-differentiable features. This is achieved by integrating an auxiliary feature function derived from Poisson's equation, which allows for precise control over the location of discontinuities. SharpNet enables joint optimization of feature locations and network parameters, outperforming existing methods in tasks like 2D problem-solving and 3D CAD reconstruction by preserving sharp edges and corners without blurring them. AI
IMPACT Introduces a new architectural approach for MLPs to better handle functions with sharp features, potentially improving performance in geometric reconstruction and other applications.