Introducing Capsa - a general framework for uncertainty-aware machine learning models
We are delighted to announce the launch of the Capsa Private Beta Program. The program began on July 1st with the selection of five participants from an extensive waitlist. These partners represent some of the most innovative, visionary, and groundbreaking organizations in the machine learning space chosen for their expertise, ongoing active projects, and commitment to deployment.
Capsa is a software framework designed to infuse arbitrary machine learning models with the ability to quantify different types of uncertainty. Several methods packaged as easy-to-use "wrappers" can be used to convert existing models into uncertainty-aware variants able to provide uncertainty estimates alongside every output. The procedure can be performed at any stage of the machine learning development process (e.g., annotation, pre-training, training, fine-tuning, inference). Capsa quantifies three main types of uncertainty: aleatoric uncertainty (due to randomness and irreducible noise), epistemic uncertainty (due to lack of knowledge), and vacuitic uncertainty (due to latent feature imbalance). The framework allows users to assess model safety and robustness.
The Capsa framework allows users to:
The framework is the culmination of over 7 years research and includes both published and proprietary algorithms developed by Themis AI. These techniques have been tested and evaluated for multiple applications including autonomous driving, face detection, drug discovery, object detection, depth estimation, text encoding, language modeling and image generation.
Capsa is currently exclusively accessible to program participants. A wider release will be announced in the future.
