Attribution Prior

Attribution priors operate by adding a regularization term to the loss function that penalizes undesirable attribution patterns. The model computes feature attributions during training and applies a penalty when those attributions diverge from expected behavior.

• Penalizes large gradients on irrelevant features
• Encourages smooth attribution maps in image models
• Can enforce sparsity in feature selection
• Operates on the input gradient or integrated gradient directly

Unlike generic regularizers such as L1 or L2 weight decay, attribution priors inject expert knowledge about what the model should consider important. This transforms opaque black-box optimization into a guided process aligned with scientific understanding.

• Radiologists can specify that diagnoses should ignore background pixels
• Financial models can be constrained to focus on known risk factors
• Climate models can enforce adherence to physical conservation laws
• Genomics models can prioritize coding regions over non-coding DNA

Modern attribution priors often use Expected Gradients as the attribution method within the regularization loop. This avoids the baseline selection problem inherent in Integrated Gradients by averaging over multiple background samples from the training distribution.

• Removes sensitivity to arbitrary baseline choice
• Provides more stable attribution estimates during training
• Enables priors that reference the data distribution rather than a single reference point
• Computationally efficient with batched background sampling

Models trained with attribution priors consistently demonstrate improved out-of-distribution generalization and resistance to spurious correlations. By forcing the model to use the right features for the right reasons, the learned representations become more causally grounded.

• Reduces reliance on dataset-specific shortcuts
• Improves performance on shifted test distributions
• Creates more interpretable decision boundaries
• Validated in medical imaging, NLP, and scientific ML domains

Attribution priors can be formulated as faithfulness constraints that require the model's explanations to accurately reflect its true decision process. This closes the gap between post-hoc rationalization and genuine model behavior.

• Penalizes explanations that don't match perturbation-based importance
• Enforces completeness so attributions sum to the prediction difference
• Can incorporate counterfactual consistency requirements
• Bridges the explainability and training objectives into a unified framework

Attribution priors are model-agnostic and can be applied to any differentiable architecture including CNNs, transformers, and graph neural networks. The only requirement is the ability to compute input gradients or integrated gradients during the forward pass.

• Compatible with PyTorch and TensorFlow autograd systems
• Works with convolutional, recurrent, and attention-based architectures
• Can regularize specific layers rather than only input features
• Extends to graph models via edge and node attribution penalties

A quantitative evaluation criterion that measures how accurately an attribution map reflects the true importance of features by perturbing the most important inputs and measuring the prediction drop.

• Mechanism: Iteratively removes or alters features ranked highest by an attribution method and records the change in model output.
• Core Assumption: A faithful explanation will cause a steep, monotonic decline in performance as critical features are corrupted.
• Relevance: The primary metric for validating the effectiveness of an Attribution Prior. A successful prior should produce models whose attributions score higher on faithfulness tests.