Stanford Seminar - Lessons from evaluating and debugging healthcare AI in deployment

Описание: James Zou, Stanford University

May 25, 2022

Dive into ai in healthcare and translating trustworthy AI from research into healthcare deployment is a major - and exciting challenge. I will discuss insights that we learned from conducting the first real-time AI trials at Stanford and analyzing data from 100 FDA-approved medical AI systems. We will explore challenges and new opportunities in each step of translation:
1. Data curation, quantifying how different data contribute to model’s success or biases
2. Model testing and monitoring, continuous real-time testing and explaining model’s mistakes
3. Human-AI interactions, designing AI for optimizing clinician’s performance.

For more information about Stanford’s Artificial Intelligence professional and graduate programs, visit: https://stanford.io/ai

0:00 Introduction
2:57 Al make clinical trials more efficient
8:27 Why did the Derm Al performance crater?
14:50 Language model captures ethnic stereotypes
16:56 Two Muslims walked into...
19:50 Data used to train dermatology Al
28:56 Data Shapley Value
29:07 Dermatology classification
30:55 Shapley value identifies mis-annotations
32:23 Data Shapley improves fairness
33:24 Auditing ML data w/ data Shapley
36:20 Understanding what the network is doing
39:19 Sparse neurons responsible for prediction
42:04 Neuron Shapley identifies dataset bias
44:26 Model repair by removing bias neurons
46:30 Why did the model make this mistake?
49:50 Conceptual explanation of mistakes Mistakes made by the model
51:44 Natural language model editing reduces bias
52:01 Takeaways: challenge shifts from model training to evaluation and monitoring

#artificialintelligence