Can I see which facial regions the AI used to come to a result?

There are several methods, such as GradCAM, that aim to highlight which facial regions contributed most to an AI’s decision. However, these approaches have limitations. The results can be inconsistent, as backtracking through the network may amplify noise, and the highlighted regions are often difficult to interpret in terms of facial shape (for example, the nose frequently appears as the main region).

To address these limitations, we are developing Face2HPO to provide a more robust and interpretable mapping of facial features to phenotypic terms.

If you would like a GradCAM-style visualization for a specific case, we can generate one for you. Please contact us directly to request it.