analiza aiqbench
This research addresses a critical bottleneck in high-dimensional variational autoencoders by reformulating latent variables using hyperspherical coordinates to mitigate volume growth issues. The proposed method significantly enhances out-of-distribution detection capabilities for both unconditional and conditional anomaly scenarios across complex real-world datasets like Mars Rover imagery and standard benchmarks such as CIFAR-10. By improving the expressiveness of approximate posteriors, this approach offers a robust advancement in machine learning model reliability without relying on external hardware or non-AI features.
VAE with Hyperspherical Coordinates: Improving Anomaly Detection from Hypervolume-Compressed Latent Space
arXiv:2601.18823v4 Announce Type: replace Abstract: Variational autoencoders (VAE) encode data into lower-dimensional latent vectors before decoding those vectors back to data. Once trained, one can hope to detect out-of-distribution (abnormal) latent vectors, but several issues arise when the latent space is high dimensional. This includes an exponential growth of the hypervolume with the dimension, which severely affects the generative capacity of the VAE. In this paper, we draw insights from high dimensional statistics: in these regimes, the latent vectors of a standard VAE are distributed on the `equators' of a hypersphere, challenging the detection of anomalies. We propose to formulate the latent variables of a VAE using hyperspherical coordinates, which allows compressing the latent vectors towards a given direction on the hypersphere, thereby allowing for a more expressive approximate posterior. We show that this improves both the fully unconditional-OOD and conditional-OOD anomaly detection ability of the VAE, achieving the best performance on the datasets we considered, outperforming existing methods. For the unconditional-OOD and conditional-OOD modalities, respectively, these are: i) detecting unusual landscape from the Mars Rover camera and unusual Galaxies from ground based imagery (complex, real world datasets); ii) standard benchmarks like Cifar10 and subsets of ImageNet as the in-distribution (ID) class.
Abstract:Variational autoencoders (VAE) encode data into lower-dimensional latent vectors before decoding those vectors back to data. Once trained, one can hope to detect out-of-distribution (abnormal) latent vectors, but several issues arise when the latent space is high dimensional. This includes an exponential growth of the hypervolume with the dimension, which severely affects the generative capacity of the VAE. In this paper, we draw insights from high dimensional statistics: in these regimes, the latent vectors of a standard VAE are distributed on the `equators' of a hypersphere, challenging the detection of anomalies. We propose to formulate the latent variables of a VAE using hyperspherical coordinates, which allows compressing the latent vectors towards a given direction on the hypersphere, thereby allowing for a more expressive approximate posterior. We show that this improves both the fully unconditional-OOD and conditional-OOD anomaly detection ability of the VAE, achieving the best performance on the datasets we considered, outperforming existing methods. For the unconditional-OOD and conditional-OOD modalities, respectively, these are: i) detecting unusual landscape from the Mars Rover camera and unusual Galaxies from ground based imagery (complex, real world datasets); ii) standard benchmarks like Cifar10 and subsets of ImageNet as the in-distribution (ID) class.
Machine Learning (cs.LG)
arXiv:2601.18823 [cs.LG]
(or arXiv:2601.18823v4 [cs.LG] for this version)
Text integral sub licența deschisă a publicației · 208 cuvinte
Subiecte conexe
High-Dimensional Latents Should Be Diagnosed Through Phase Structure
arXiv · 3 iun. 2026
[AINews] not much happened today
Latent Space · 3 oct. 2026
Vertical Advantage: Transforming Industries with Lakebase and Agentic AI
Databricks · 27 aug. 2026
Article: Multi-Agent AI for Production Security Operations: an A2A and MCP Architecture in a 5G Core
InfoQ · 23 iul. 2026