Hyperprior

Hyperprior Research is a London-based R&D lab specializing in complex systems and robust probabilistic modeling.

Hyperprior Research operates at the intersection of complex systems and advanced machine learning from its base in London. The core technology (the Hyperprior model) revolutionized neural image compression by introducing a hierarchical bottleneck: a primary autoencoder transforms data into latents, while a second hyper-network captures the spatial dependencies within those latents to guide the entropy coder. This two-part system allows for context-aware bit allocation, enabling models to outperform traditional codecs like JPEG and BPG by focusing bits on high-entropy regions. By treating hyperparameters as learnable distributions rather than fixed constants, the framework achieves superior rate-distortion performance across visual and hyperspectral data sets.

https://hyperprior.net