Dissertation on feature construction in vision models

• Read PDF

How inhibitory weights enable the construction of selective features.

• Read PDF

Image seeded feature visualizations reveal 'what' a feature sees in a given image

• Paper

Activation maximization isn't just about visualizing features, it can also tell us about what the network can/can't represent.

• Paper

Automated circuit extraction in CNNs

• Paper

A tool for Visualizing Interpretable Subgraphs in CNNs

• Read PDF

Feeding an autoencoders outputs as inputs lets you visualize the fixed-points of their embedding space.

A paper testing which model objectives/architectures account for human representational geometry of letter perception. With Daniel Janini

• Paper