Invariant Transformation and Resampling based Epistemic-Uncertainty Reduction

Researchers propose a new AI inference method that uses invariant transformations and resampling to reduce epistemic uncertainty and improve model accuracy. The approach involves applying multiple transformed versions of an input to a trained AI model and aggregating the outputs for more reliable results.

• →AI models suffer from inference errors due to aleatoric and epistemic uncertainties even when well-trained.
• →Multiple samples based on invariant transformations of inputs can show partial independence in inference errors.
• →The proposed resampling method aggregates outputs from multiple transformed inputs to improve accuracy.
• →This approach offers a strategy for balancing model size and performance requirements.
• →The method can be applied to existing trained AI models without requiring retraining.