A Comprehensive Survey on Evidential Deep Learning and its Applications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models require reliable uncertainty estimation for safe deployment in high-risk applications.
• Existing methods like deep ensembling and Bayesian neural networks are computationally expensive.
• Evidential Deep Learning (EDL) provides an efficient alternative for uncertainty estimation.

Purpose of the Study:

• To provide a comprehensive survey of Evidential Deep Learning (EDL).
• To introduce EDL to readers without prior knowledge.
• To cover theoretical foundations, advancements, applications, and future directions of EDL.

Main Methods:

• Review of subjective logic theory as the foundation of EDL.
• Exploration of EDL advancements: evidence collection, OOD sample utilization, training strategies, and evidential regression.
• Discussion of EDL applications across diverse machine learning tasks.

Main Results:

• EDL enables high-quality uncertainty estimation with minimal computational overhead.
• EDL offers a distinct approach compared to other uncertainty estimation frameworks.
• EDL has demonstrated broad applicability in various machine learning paradigms.

Conclusions:

• EDL is a promising paradigm for efficient and reliable uncertainty estimation.
• Further research in EDL can enhance its performance and adoption.
• EDL has the potential to significantly impact fields requiring robust AI decision-making.