Been Kim: Interpretable Machine Learning
The Gradient: Perspectives on AI
00:00
How Different Architectures Influence Different Concepts?
We wrote a paper on concept bottleneck models. This is one of the patterns in Brazil, where we insert those unknown concepts that we want to be able to have access to and estimate importance of them. And we show pretty, you know, actually medical data in the paper that this controllability improves the accuracy too. But I liked the way that you talked about how different in inductive biases and how different architectures influence different concepts. In fact, I would even argue like CNN, we don't really understand in theory, yes, translation and variance should happen.
Transcript
Play full episode
