Scattering Network

Scattering networks are a class of convolutional neural networks employing predefined filters, typically wavelets, to extract robust and interpretable features from various data types, including images, signals, and graphs. Current research focuses on developing novel scattering network architectures, such as those incorporating transformers or leveraging multiscale basis dictionaries on simplicial complexes, to improve performance and address limitations like slow energy decay or oversmoothing in traditional designs. These networks find applications across diverse fields, from medical image analysis and cosmological inference to material science and knowledge graph analysis, offering advantages in terms of computational efficiency, feature interpretability, and robustness to noise and data variations.