Publikation
Learning to Intervene on Concept Bottlenecks
David Steinmann; Wolfgang Stammer; Felix Friedrich; Kristian Kersting
In: Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27, 2024. International Conference on Machine Learning (ICML), OpenReview.net, 2024.
Zusammenfassung
While deep learning models often lack inter-
pretability, concept bottleneck models (CBMs)
provide inherent explanations via their concept
representations. Moreover, they allow users to per-
form interventional interactions on these concepts
by updating the concept values and thus correct-
ing the predictive output of the model. Up to this
point, these interventions were typically applied
to the model just once and then discarded. To
rectify this, we present concept bottleneck mem-
ory models (CB2Ms), which keep a memory of
past interventions. Specifically, CB2Ms leverage
a two-fold memory to generalize interventions to
appropriate novel situations, enabling the model
to identify errors and reapply previous interven-
tions. This way, a CB2M learns to automatically
improve model performance from a few initially
obtained interventions. If no prior human inter-
ventions are available, a CB2M can detect poten-
tial mistakes of the CBM bottleneck and request
targeted interventions. Our experimental evalua-
tions on challenging scenarios like handling dis-
tribution shifts and confounded data demonstrate
that CB2Ms are able to successfully generalize in-
terventions to unseen data and can indeed identify
wrongly inferred concepts. Hence, CB2Ms are a
valuable tool for users to provide interactive feed-
back on CBMs, by guiding a user’s interaction
and requiring fewer interventions.