FiD-GP is a compression framework that incorporates compact guided weight matrices to project the weight uncertainty of a neural network into a low-dimensional subspace. It enhances expressiveness through a regularization flow dictionary and spectral normalization, and aligns the guided subspace with the feature-gradient geometry through a numerically stable projection mechanism. Using FiD-GP's prediction framework, we can design a single-pass projection for Out-of-Distribution (OoD) detection. It improves uncertainty estimation performance compared to SVGP-based baselines on regression, image classification, semantic segmentation, and OoD detection benchmarks, satisfies tight spectral residual bounds with theoretically guaranteed OoD detection, and significantly compresses the storage requirements of neural networks.
