BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Learning to Receive Help: Intervention-Aware Concept Embedding Mod
 els - Mateo Espinosa Zarlenga (University of Cambridge)
DTSTART:20240213T130000Z
DTEND:20240213T140000Z
UID:TALK207517@talks.cam.ac.uk
CONTACT:Mateja Jamnik
DESCRIPTION:Concept Bottleneck Models (CBMs) tackle the opacity of neural 
 architectures by constructing and explaining their predictions using a set
  of high-level concepts. A special property of these models is that they p
 ermit concept interventions\, wherein users can correct mispredicted conce
 pts and thus improve the model's performance. Recent work\, however\, has 
 shown that intervention efficacy can be highly dependent on the order in w
 hich concepts are intervened on and on the model's architecture and traini
 ng hyperparameters. In this talk\, I will argue that this is rooted in a C
 BM's lack of train-time incentives for the model to be appropriately recep
 tive to concept interventions. To address this\, I will propose Interventi
 on-aware Concept Embedding models (IntCEMs)\, a novel CBM-based architectu
 re and training paradigm that improves a model's receptiveness to test-tim
 e interventions. Our model learns a concept intervention policy in an end-
 to-end fashion from where it can sample meaningful intervention trajectori
 es at train-time. This conditions IntCEMs to effectively select and receiv
 e concept interventions when deployed at test-time. Our experiments show t
 hat IntCEMs significantly outperform state-of-the-art concept-interpretabl
 e models when provided with test-time concept interventions\, demonstratin
 g the effectiveness of our approach.\n\n"You can also join us on Zoom":htt
 ps://cam-ac-uk.zoom.us/j/92041617729
LOCATION:Lecture Theatre 2\, Computer Laboratory\, William Gates Building
END:VEVENT
END:VCALENDAR