Co-authors: Antonio Mastropietro, Marco Nurisso and Francesco Vaccarino
📄 Read the full paper: openreview.net/pdf?id=VY74p...
If you're at #ICML2025, catch Antonio at his poster!
📍 West Exhibition Hall B2-B3 W-201
🗓️ Wed 16 Jul | ⏰ 4:30–7 p.m.

#ICML2025 #DeepLearning #Interpretability #AI

8️⃣

These vector fields allow us to compare attributes in terms of invariance quantification ! Roughly, the more disordered the extracted vector field, the more invariant the model is to the considered attribute.
We validate this finding with a finetunings protocol.

7️⃣

At the microscale for continuous attributes, we instead derive vector fields with the help of generative models to leverage directional information only and avoid the problem of incomparable scales (How would you compare ageing by two years vs having blonder hair ?).

6️⃣

At the macroscale and for binary attributes, we compare distribution of distances in the embedding space.
Spoiler: Baldness and gender affect the model way more than, say, smiling.

5️⃣

However, face recognition embeddings have a hierarchical structure inherited during the contrastive learning and we distinguish two scales:
🔹 Micro-scale: How one person's photos vary.
🔹 Macro-scale: How identities, groups of photos are distributed.

4️⃣

These questions fall under sensitivity analysis — a way to study which features really impact learned representations.

We propose a method using geometric tools that respect how embeddings are structured.
This helps us measure invariance — how resistant a model is to specific input changes.

3️⃣

Our focus: How do face recognition models react to what-if scenarios?
E.g.:
🔹 What if she had eyeglasses?
🔹 What if he were slightly older ?
🔹 What if his hair were lighter ?
We want to quantify this through looking at the embedding space - the internal representation space of the model.

2️⃣