ML708-Explainability-2

Which of the following are the ingredients for an interpretability method?

• Model

• Data

• Humans

• Task

• All of these

Concept activation vector is a local interpretability method that quantifies the importance of a concept towards a trained deep neural network.

• True

• False

One can validate the CAVs using

• By sorting the dataset using cosine similarity

• Calculating TCAV on complete dataset

• By comparing with Saliency Maps

Concept Activation Vectors (CAVs) do not require positive/negative dataset

• True

• False

Unintended entanglement of concepts can be solved using

• Concept Activation Vector (TAV)

• Using concepts from CLIP latent space

• Using small number of carefully chosen concept-annotated images

Advantage of projection and Inverse Projection in Clip-based CounTEX

• Can be learned on small dataset with annotations

• Can be learned on any other dataset with annotations

• Can be learned on any other dataset without annotations

• correct prediction

• incorrect prediction