Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causally Reliable Concept Bottleneck Models

Giovanni De Felice, Arianna Casanova Flores, Francesco De Santis, Silvia Santini, Johannes Schneider, Pietro Barbiero, Alberto Termine

TL;DR

The paper addresses the fragility of purely associative concept bottleneck models by introducing Causally reliable Concept Bottleneck Models (C$^2$BMs), which impose a causal bottleneck over concepts guided by a structural causal model. It presents a fully automated pipeline to discover concepts and causal graphs from data and background knowledge, and to train an exogenous-embedding encoder and a hypernetwork that parameterizes causal mechanisms. Empirical results across synthetic and real datasets show that C$^2$BMs maintain competitive task accuracy while achieving improved causal reliability, better responsiveness to interventions, and enhanced debiasing and fairness properties. This work offers a principled path toward interpretable, causally grounded AI that supports interventions and fairness in complex decision tasks.

Abstract

Concept-based models are an emerging paradigm in deep learning that constrains the inference process to operate through human-interpretable variables, facilitating explainability and human interaction. However, these architectures, on par with popular opaque neural models, fail to account for the true causal mechanisms underlying the target phenomena represented in the data. This hampers their ability to support causal reasoning tasks, limits out-of-distribution generalization, and hinders the implementation of fairness constraints. To overcome these issues, we propose Causally reliable Concept Bottleneck Models (C$^2$BMs), a class of concept-based architectures that enforce reasoning through a bottleneck of concepts structured according to a model of the real-world causal mechanisms. We also introduce a pipeline to automatically learn this structure from observational data and unstructured background knowledge (e.g., scientific literature). Experimental evidence suggests that C$^2$BMs are more interpretable, causally reliable, and improve responsiveness to interventions w.r.t. standard opaque and concept-based models, while maintaining their accuracy.

Causally Reliable Concept Bottleneck Models

TL;DR

The paper addresses the fragility of purely associative concept bottleneck models by introducing Causally reliable Concept Bottleneck Models (CBMs), which impose a causal bottleneck over concepts guided by a structural causal model. It presents a fully automated pipeline to discover concepts and causal graphs from data and background knowledge, and to train an exogenous-embedding encoder and a hypernetwork that parameterizes causal mechanisms. Empirical results across synthetic and real datasets show that CBMs maintain competitive task accuracy while achieving improved causal reliability, better responsiveness to interventions, and enhanced debiasing and fairness properties. This work offers a principled path toward interpretable, causally grounded AI that supports interventions and fairness in complex decision tasks.

Abstract

Concept-based models are an emerging paradigm in deep learning that constrains the inference process to operate through human-interpretable variables, facilitating explainability and human interaction. However, these architectures, on par with popular opaque neural models, fail to account for the true causal mechanisms underlying the target phenomena represented in the data. This hampers their ability to support causal reasoning tasks, limits out-of-distribution generalization, and hinders the implementation of fairness constraints. To overcome these issues, we propose Causally reliable Concept Bottleneck Models (CBMs), a class of concept-based architectures that enforce reasoning through a bottleneck of concepts structured according to a model of the real-world causal mechanisms. We also introduce a pipeline to automatically learn this structure from observational data and unstructured background knowledge (e.g., scientific literature). Experimental evidence suggests that CBMs are more interpretable, causally reliable, and improve responsiveness to interventions w.r.t. standard opaque and concept-based models, while maintaining their accuracy.

Paper Structure

This paper contains 57 sections, 1 theorem, 7 equations, 18 figures, 13 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Concept Bottleneck Models.
  4. Causal Reliability.
  5. Structural Causal Models.
  6. Related works
  7. Method
  8. Causally reliable Concept Bottleneck Models
  9. Model instantiation
  10. Concept discovery and labeling
  11. Causal graph discovery
  12. Structural equations and model training
  13. Model training.
  14. Experimental evaluations
  15. Task accuracy and causal reliability
  16. ...and 42 more sections

Key Result

Theorem D.1

C$^2$BM is a universal approximator regardless of the underlying causal graph.

Figures (18)

  • Figure 1: Causally reliable Concept Bottleneck Models (C$^2$BMs) enforce reasoning through a "Causal Bottleneck" aligned with a model of real-world causal mechanisms obtained from data and background knowledge.
  • Figure 2: Overview of the C$^2$BM fully automated pipeline. The pipeline consists of three key blocks: (i) discovery and labeling of the relevant variables $\mathcal{V}$ from background knowledge; (ii) discovery of the causal graph by integrating data and background knowledge; (iii) a C$^2$BM model, comprising a neural encoder and an adaptively parametrized SCM. Once the model is trained, it can support queries about any endogenous variable (e.g., predicting dyspea).
  • Figure 3: Probabilistic graphical model of C$^2$BM inference.
  • Figure 4: Label accuracy (%) on downstream variables (task included) after intervening on concepts up to progressively deeper levels in the graph hierarchy. Summit plots show the difference of C$^2$BM's accuracy w.r.t. the best-performing baseline. Uncertainties represent $2$ sample mean $\sigma$ across 5 runs.
  • Figure 5: Biased ColorMNIST dataset. Task accuracy on Parity after ground-truth interventions.
  • ...and 13 more figures

Theorems & Definitions (4)

  • Remark 4.1
  • Remark 4.5
  • Theorem D.1
  • proof