Conceptual Contrastive Edits in Textual and Vision-Language Retrieval
This addresses interpretability challenges for researchers and practitioners working with textual and vision-language retrieval models, though it appears incremental as it builds on existing contrastive editing approaches.
The paper tackles the problem of model-agnostic interpretability for complex deep learning models by using post-hoc conceptual contrastive edits to expose patterns and biases in retrieval model representations, resulting in a systematic method for designing interventions and a novel metric to evaluate their per-word impact.
As deep learning models grow in complexity, achieving model-agnostic interpretability becomes increasingly vital. In this work, we employ post-hoc conceptual contrastive edits to expose noteworthy patterns and biases imprinted in representations of retrieval models. We systematically design optimal and controllable contrastive interventions targeting various parts of speech, and effectively apply them to explain both linguistic and visiolinguistic pre-trained models in a black-box manner. Additionally, we introduce a novel metric to assess the per-word impact of contrastive interventions on model outcomes, providing a comprehensive evaluation of each intervention's effectiveness.