Prototype-Based Explanations
A vision model labels a photograph “chapel.” Instead of returning a heatmap, it shows training images with similar internal representations.
These prototypes answer a familiar question: what known cases does this look like? Criticisms show nearby cases the prototype set fails to represent — instances close in feature space to a prototype that the prototype set does not explain well.
Compare Prototypes and Criticisms
Step through nearby training examples in feature space. Similarity provides the ranking; the criticism shows what the current prototype set fails to represent.
Chapel A — Duke Chapel in winter, Durham NC. Representative prototype. Wikipedia
Similarity in feature space
Prototypes are selected using distance between internal feature vectors, often Euclidean or cosine distance. MMD-Critic instead chooses a set that reduces the discrepancy between the prototype distribution and the full data distribution.
MMD-Critic — Algorithm Overview
Step 1 — Choose counts
Decide how many prototypes m and criticisms c you want. These are the only free parameters.
Prototypes summarize the data distribution; criticisms highlight where prototypes fall short. Choosing m and c is a design decision — more prototypes give a richer summary but are harder to inspect.
m — Prototypes
Cover the typical cases. Start with 3–10% of dataset size.
c — Criticisms
Flag edge cases. Often set to the same count as m.
Choose counts
Decide how many prototypes m and criticisms c you want. Thes…
Select prototypes
Run a greedy search over the data. At each step, pick the po…
Select criticisms
Compute the witness function for every non-prototype point. …
Return results
The final prototype set summarizes the dataset. The final cr…
Examples are concrete, which helps in medicine, fraud review, and quality control. Their familiarity can also over-persuade. A close embedding does not guarantee the model matched on the property a person notices.
Show similarity scores, include contrastive cases, and check whether the selected examples cover important subgroups.
Prototypes and criticisms are always actual instances from the data
Unlike synthetic explanations, prototypes and criticisms are real training examples — nothing is generated or interpolated. This grounds the explanation in the actual data the model learned from, but it also means the explanation is limited by what exists in that dataset.
Why should a prototype explanation include a criticism?