Quantitative Testing with Concept Activation Vectors (TCAV) - 2018

Explore the concept of Quantitative Testing with Concept Activation Vectors (TCAV) in this 56-minute talk by Been Kim from Google, presented at the Center for Language & Speech Processing (CLSP) at JHU in 2018. Dive into the challenges of interpreting deep learning models and discover how CAVs provide a novel approach to understanding neural networks' internal states using human-friendly concepts. Learn about the TCAV technique, which uses directional derivatives to quantify the importance of user-defined concepts in classification results. Follow the speaker's journey through image classification examples, including a medical application for diabetic retinopathy. Gain insights into the potential of CAVs to empower humans in their interaction with machine learning systems and explore the implications for building more interpretable AI models.

Intro
Types of interpretability methods
Goal
our work TCAV: Testing with Concept Activation Vectors
Defining concept activation vector (CAV)
TCAV core idea: Use CAV to get prediction sensitivity
TCAV is generalization of saliency maps for concepts
recap: Testing with Concept Activation Vectors
Sanity check experiment setup
Human subject experiment: Can saliency maps communicate the same information?
Googlenet
Inception V3
TCAV for Diabetic Retinopathy
Summary: Testing with Concept Activation Vectors
Questions?