Do Concept Bottleneck Models Learn as Intended?
Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel pith:LNX4ZISJrecord.jsonopen to challenge →
read the original abstract
Concept bottleneck models map from raw inputs to concepts, and then from concepts to targets. Such models aim to incorporate pre-specified, high-level concepts into the learning procedure, and have been motivated to meet three desiderata: interpretability, predictability, and intervenability. However, we find that concept bottleneck models struggle to meet these goals. Using post hoc interpretability methods, we demonstrate that concepts do not correspond to anything semantically meaningful in input space, thus calling into question the usefulness of concept bottleneck models in their current form.
This paper has not been read by Pith yet.
Forward citations
Cited by 12 Pith papers
-
Concept Flow Models: Anchoring Concept-Based Reasoning with Hierarchical Bottlenecks
Concept Flow Models use hierarchical concept-driven decision trees to mitigate information leakage in concept bottleneck models while matching their predictive performance.
-
In Defense of Information Leakage in Concept-based Models
Concept-based models can use controlled 'benign' information leakage to remain accurate and intervenable under real-world concept incompleteness by reframing their training objective.
-
OceanCBM: A Concept Bottleneck Model for Mechanistic Interpretability in Ocean Forecasting
OceanCBM is the first concept bottleneck model for spatiotemporal ocean prediction that uses mixed supervision on physical concepts and a free concept to deliver consistent mechanistic representations for mixed layer ...
-
Caption Bottleneck Models
Caption Bottleneck Models use LMM-generated image captions as the sole input to a text classifier, creating leakage-free interpretable models that discover dataset-specific concepts without predefined lists or manual labels.
-
GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis
GRAPE augments prototype medical image classifiers with graph attention for co-occurrence, a mismatch safety check, and open-vocabulary anchoring to support incremental findings without retraining.
-
GRAPE: Graph-Augmented Prototype Explanations for Interactive Medical Image Diagnosis
GRAPE augments prototype medical image classifiers with graph attention for co-occurrence, a mismatch safety check, and open-vocabulary anchoring to support incremental addition of findings from single examples.
-
SynCB: A Synergy Concept-Based Model with Dynamic Routing Between Concepts and Complementary Neural Branches
SynCB adds a dynamic routing module and joint training to a hybrid concept-plus-neural architecture, reporting up to 3.9 pp higher accuracy than a full neural baseline and up to 6.43 pp better intervention responsiven...
-
Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations
VH-CBM uses a Gaussian process in VLM embedding space to propagate sparse human annotations and improve concept accuracy and calibration over pure VLM-guided concept bottleneck models.
-
Sparse Concept Anchoring for Interpretable and Controllable Neural Representations
Sparse Concept Anchoring biases neural latent spaces toward targeted concepts using under 0.1% labels per concept, enabling reversible steering via projection and permanent removal via weight ablation with minimal sid...
-
Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers
ViSAE supplies a 64K-image probing suite with 16K concepts, top-down/bottom-up circuit algorithms, and editing methods that raise WaterBirds worst-group accuracy by 48.2% over baselines.
-
CHiQPM: Calibrated Hierarchical Interpretable Image Classification
CHiQPM is a hierarchical interpretable image classifier that maintains 99% of non-interpretable model accuracy while supplying contrastive global explanations, human-like hierarchical paths, and calibrated interpretab...
-
If Concept Bottlenecks are the Question, are Foundation Models the Answer?
Empirical tests of VLM-CBMs show VLM supervision differs from expert annotations depending on task and that concept accuracy correlates weakly with quality metrics.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.