How Well do Feature Visualizations Support Causal Understanding of CNN Activations?

Institute Homepage

Institute Homepage Sign In

Back

Research Overview

Comparative Vision Science

Does CLIP's Generalization Performance Mainly Stem from High Train-Test Similarity?

In Search of Forgotten Domain Generalization

Partial Success in Closing the Gap between Human and Machine Vision

Foundations of Machine Learning

Cross-Entropy Is All You Need To Invert the Data Generating Process

Contrastive Learning Inverts the Data Generating Process

Interaction Asymmetry: A General Principle for Learning Composable Abstractions

Mechanistic Interpretability

Measuring Per-Unit Interpretability at Scale Without Humans

Exemplary Natural Images explain CNN Activations better than State-Of-The-Art Feature Visualizations

How Well do Feature Visualizations Support Causal Understanding of CNN Activations?

Robust Machine Learning

How Well do Feature Visualizations Support Causal Understanding of CNN Activations?

We introduce a new psychophysical task to measure how well visualizations help humans predict the causal effects of interventions on unit activations, finding that widely-used feature visualizations provide no significant advantage over simpler alternatives like dataset samples in fostering causal understanding~[