Concept Bottleneck Models
A medical model can first predict clinically meaningful concepts such as lesion shape, border irregularity, and color variation, then use those concepts to predict diagnosis.
The concept layer creates a place where a clinician can inspect or correct the model's intermediate state.
Predict through concepts
A concept bottleneck factors prediction into:
predicts concept vector from input . predicts outcome from those concepts. Intervention replaces one predicted concept with a trusted value.
Correct a Clinical Concept Before Diagnosis
A skin-lesion model first predicts three dermatology concepts. Correct the border-irregularity concept and watch only the downstream diagnosis update.
Model architecture
Only the path through c reaches h — g and h cannot see x directly, so an intervention on c is the only way to change y without retraining.
Melanoma risk
75%
risk = .08
+ .34(asymmetry)
+ .39(border)
+ .19(color)
Intervention effect
0.0 points
The diagram traces : predicts the three concept bars from the image, and sees only those scores on its way to the risk output — never the image itself. That handoff is the bottleneck, and the only place a human can step in.
The "Clinician corrects border to 0.35" toggle is that step-in: it swaps the model's border-irregularity estimate for a clinician-trusted 0.35, as if an expert overruled the model on that one concept. Since only reads concept scores, this recomputes the risk instantly — no rerunning , no retraining — and the diagram highlights the corrected bar and the intervention arrow so you can see that one correction's isolated effect.
Bottlenecks make reasoning inspectable, but only through the concepts you chose. An incomplete vocabulary can force the model to lose useful information or smuggle it through concept probabilities.
Evaluate concept accuracy, downstream accuracy, intervention benefit, and whether the vocabulary covers known decision factors.
What happens when an important concept is missing from the bottleneck?