← Back to Projects
Geometry-Aware Generative Modelling for Out-of-Distribution Detection
Addressing deep learning safety vulnerabilities with von Mises–Fisher (vMF) distributions.
- Developed a geometry-aware Variational Autoencoder with a spherical latent space governed by the von Mises–Fisher (vMF) directional distribution, addressing safety vulnerabilities where foundation model encoders give confidently wrong predictions under distribution shifts.
- Derived and implemented the closed-form expression for the KL divergence under a uniform hyperspherical prior, avoiding computationally expensive numerical approximations and eliminating latent sampling noise.
- Achieved a +8% to 9% AUROC improvement in near out-of-distribution evaluation tasks over baseline models, at the cost of increased false positive rates in far out-of-distribution detection.