TCAV
A concept-based model interpretability method that trains linear classifiers, called concept activation vectors, in the intermediate-layer activation space of neural networks to quantify the influence of human-interpretable concepts on model predictions.
Timeline (last 90 days)
TCAV (Testing with Concept Activation Vectors) quantifies the influence of a given concept on prediction results by training a linear classifier in the activation space of the model's intermediate layers.
Methods such as ACE and ConceptSHAP extend TCAV but retain the prerequisite of requiring access to the model's internals.
Researchers transformed the key features identified by SHAP into concept sets for TCAV (Testing with Concept Activation Vectors) to probe the multilingual XLM-R model and various language-specific encoders.
All Facts (3)
TCAV (Testing with Concept Activation Vectors) quantifies the influence of a given concept on prediction results by training a linear classifier in the activation space of the model's intermediate layers.
50%UnverifiedMethods such as ACE and ConceptSHAP extend TCAV but retain the prerequisite of requiring access to the model's internals.
50%UnverifiedResearchers transformed the key features identified by SHAP into concept sets for TCAV (Testing with Concept Activation Vectors) to probe the multilingual XLM-R model and various language-specific encoders.
50%