Search Papers | Poster Sessions | All Posters
Poster C18 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Optimizing for Possible Feature Combinations in Discriminative Vision Models
Christopher Hamblin1 (), Talia Konkle1, George Alvarez1; 1Harvard University
In this work we leverage feature visualization to probe the bounds of feature combinations in the InceptionV1 object recognition model. While our technique also yields conventional/viewable feature visualizations, we demonstrate how such optimizations can reveal contingencies between feature pairs that are difficult to infer from their activations to natural images alone. We propose a data visualization motif that is ideal for quickly assessing the relations between arbitrary feature pairs.
Keywords: Neural Networks Mechanistic Interpretability Feature Visualization CNNs