Concept Bottleneck Model

The CBM architecture enforces a strict information bottleneck by decoupling prediction into sequential stages:

• Stage 1 — Concept Encoder: Maps raw input x (e.g., a spectrogram or IQ sample) to a set of concept activation scores ĉ = g(x). Each concept corresponds to a human-defined, semantically meaningful attribute.
• Stage 2 — Task Predictor: Makes the final prediction ŷ = f(ĉ) using only the concept scores as input. The model is structurally prohibited from using raw features directly.

This forced routing through the concept layer ensures that every downstream decision can be traced back to a weighted combination of interpretable concepts.

A defining capability of CBMs is test-time concept intervention, where a domain expert can manually edit concept activations and observe how the final prediction changes:

• If a model misclassifies a signal because it incorrectly predicts the concept "modulation is QPSK," an operator can override that concept to "modulation is 16-QAM" and see the corrected output.
• This enables real-time human-AI collaboration without retraining, making CBMs uniquely suited for mission-critical RF applications where operator trust is non-negotiable.
• Intervention efficacy is measured by how accurately the final prediction reflects the corrected concept values, validating that the bottleneck is genuinely causal.

CBMs require concept-level supervision during training, which introduces a distinct data labeling burden:

• Each training sample must be annotated not only with the final task label but also with ground-truth values for every concept in the bottleneck layer.
• For RF applications, concepts might include modulation type, symbol rate, carrier frequency offset, or signal-to-noise ratio.
• This contrasts with standard end-to-end black-box models that learn latent features without human specification. The trade-off is interpretability versus annotation cost.
• Hybrid approaches use unsupervised concept discovery to reduce labeling requirements while preserving partial interpretability.

The concept bottleneck imposes a sequential information constraint that can reduce representational capacity compared to unrestricted architectures:

• Standard model: Input → Hidden layers → Output (no constraint on learned features)
• CBM: Input → Concepts → Output (forced compression through human-specified concepts)

Research shows that when the concept set is complete — meaning all information necessary for the task is captured by the chosen concepts — CBMs can match black-box accuracy. When concepts are incomplete, performance degrades.

Mitigation strategies include residual connections that allow a controlled amount of non-concept information to bypass the bottleneck, trading some interpretability for accuracy.

CBMs are particularly well-suited for explainable RF machine learning where operator trust and regulatory compliance are paramount:

• Concept examples for RF: Modulation scheme, baud rate, pulse shaping filter, center frequency offset, presence of interference, SNR estimate.
• A CBM for automatic modulation classification would first predict these intermediate signal properties, then classify the modulation based solely on those human-verifiable concepts.
• This allows a signals analyst to audit the reasoning chain: "The model classified this as QPSK because it detected a constant envelope, phase shifts at symbol boundaries, and a symbol rate of 1 MHz."
• In contested or adversarial RF environments, the ability to intervene on mispredicted concepts before final classification provides a critical safety layer.

Two critical failure modes must be addressed when designing a CBM:

• Concept incompleteness: If the chosen concept set does not capture all task-relevant information, the model cannot achieve optimal performance. For example, if "phase noise" is omitted from the concept set but is critical for emitter identification, accuracy will suffer.
• Concept leakage: If the concept encoder inadvertently encodes task-relevant information that is not captured by the specified concepts, the bottleneck is compromised. This can occur when the encoder learns to embed residual task information in the concept activations.

Detection methods include training a separate model to predict the final task from the concept scores alone — if this model underperforms the full CBM, leakage may be present.

A game-theoretic framework that assigns each input feature an importance score for a particular prediction. Based on Shapley values from cooperative game theory, SHAP fairly distributes the model output among features.

• Provides both local and global explanations
• Model-agnostic: works with any black-box model
• Satisfies efficiency, symmetry, and additivity properties
• Computationally expensive for high-dimensional inputs

A model-agnostic technique that explains individual predictions by approximating the complex model locally with an interpretable surrogate. LIME generates perturbed samples around the instance of interest and fits a simple model like a linear regressor.

• Creates locally faithful explanations
• Works on tabular, text, and image data
• Surrogate model choice affects explanation quality
• Does not guarantee global fidelity

An attribution method that assigns importance to input features by accumulating gradients along a straight-line path from a baseline input to the actual input. Satisfies the completeness axiom: attributions sum to the difference between the output at the input and the baseline.

• Requires careful baseline selection
• Implementation invariance guaranteed
• Particularly effective for deep neural networks
• Computationally more demanding than simple gradient methods

A glass-box model that combines the accuracy of gradient boosting with complete intelligibility. EBMs learn generalized additive models where each feature contributes through a learned function, making every prediction fully decomposable.

• Native interpretability without post-hoc analysis
• Feature functions can be plotted and inspected
• Competitive accuracy with black-box models
• Supports pairwise interaction terms while maintaining explainability

A causal explanation method that identifies the minimal change to an input instance required to alter a model's prediction. Unlike feature attribution, counterfactuals answer 'what if' questions directly.

• Generates actionable recourse: 'Change X to get outcome Y'
• Naturally aligns with human reasoning patterns
• Multiple valid counterfactuals may exist
• Used extensively in credit scoring and hiring decisions

A subfield of AI safety that reverse-engineers neural network internals into human-understandable algorithms. Treats models as scientific objects of study, identifying circuits and mechanisms rather than just feature importance.

• Seeks to understand how models compute, not just what they attend to
• Identifies polysemantic neurons and superposition
• Complements CBM approaches by validating learned concepts
• Critical for alignment and safety verification