Gastroenterology and Artificial Intelligence: 2nd Annual Artificial Intelligence Summit | October 2020-The Algorithms An Alphabet Soup

The Algorithms An Alphabet Soup_Bagci

Pdf Summary

The document discusses the drawbacks of current AI algorithms and the need for Explainable AI (XAI) and trust in AI systems. It highlights the failures of non-robust algorithms in AI and the importance of having algorithms that are explainable and not fragile. The concept of interpretable deep learning (DL) is introduced, which refers to the ability to interpret DL models in a domain-specific manner. Current approaches for XAI are discussed, including altering pixels to observe the impact on predictions and the use of attention mechanisms. However, these approaches have drawbacks such as being post-hoc and lacking reasoning or visual attributes.<br /><br />The document proposes the idea of inherently explainable algorithms that are robust and not fragile. It presents an example of a human-in-the-loop XAI system that combines human expertise with AI to achieve better results. The example involves the integration of multi-parametric image data and the use of eye tracking to identify areas at risk for missed diagnosis. The document suggests that eye tracking can also be used for fast segmentation and ground truth labeling.<br /><br />Overall, the document emphasizes the need for transparency, trust, and understanding in AI algorithms. It concludes that XAI can increase transparency and build trust by providing interpretations and understanding of how DL models work. It also mentions the importance of fair and causal models in AI.

Keywords

AI algorithms

Explainable AI

trust in AI systems

interpretable deep learning

XAI approaches

altering pixels

attention mechanisms

inherently explainable algorithms

human-in-the-loop XAI

transparency in AI