Gradient-based explanations can reveal model structure.
problem Tension between secrecy and model explanations.
method Algorithm to learn a two-layer ReLU network using gradient queries.
result The number of gradient queries is nearly optimal and independent of model size.
Proposes a new method for better explaining neural network decisions.
problem Challenges in explaining neural network decisions due to base-point choice.
method Introduces tangentially aligned integrated gradients to maximize explanation tangential alignment.
result Optimal base-point maximizes explanation tangential alignment, leading to more accurate interpretations.
The paper explores additive explanations for non-additive models, finding that non-additive methods are more accurate.
problem Explaining non-additive models using additive methods.
method Four explanation methods: partial dependence, Shapley, distilled, and gradient-based.
result Non-additive explanations are more accurate than distilled additive explanations.
Differentially private algorithms protect model explanations from leaking training data.
problem Model explanations can leak training data, compromising privacy.
method Adaptive differentially private gradient descent algorithm to produce accurate, private explanations.
result Privacy amplification and reduction of overall privacy loss on explanation data.
Paper compares two local explanation methods for machine learning models.
problem Comparing two local explanation methods for machine learning models.
method Integrated Gradients and Baseline Shapley methods.
result Additional insights on comparative behavior for tabular data and neural networks.
Gradient-based explanations correlate with Android malware classifier robustness.
problem Evasion attacks on Android malware classifiers using sparse perturbations.
method Investigated gradient-based attribution methods for explaining classifier decisions and their evenness, proposing metrics to assess adversarial robustness.
result Gradient-based explanations, especially Integrated Gradients, correlate with adversarial robustness of malware classifiers.
Study evaluates relevance metrics for similarity-based model explanations.
problem Providing understandable explanations for complex model predictions.
method Evaluated three relevance metrics using three tests.
result Cosine similarity of gradients performs best for explanations.
We generate counterfactual explanations for tree-based boosting ensembles.
problem Understanding how tree-based models make predictions.
method Extending a method for random forests to GBDTs, accounting for tree sequential dependency and negative gradients.
result A method to generate counterfactual explanations for GBDTs.
The paper introduces a method to learn models with built-in explanations.
problem Lack of interpretability in deep learning models.
method Formalizes learning with explanation constraints and provides a learning theoretic framework.
result Models that satisfy these constraints have reduced Rademacher complexities, improving their performance.
Gradient-based methods improve understanding of deep learning survival models.
problem Limited interpretability of deep learning survival models hinders their adoption.
method Gradient-based explanation methods tailored to survival neural networks.
result Gradient-based methods capture feature effects and temporal dynamics.
Study examines explainable machine learning for monotonic models, finding Integrated gradients better for strong monotonicity.
problem Applying explainable machine learning to science-informed models.
method Proposed axioms for monotonicity, tested Shapley value and Integrated gradients methods.
result Integrated gradients provides better explanations for strong monotonicity.
A technique scales symbolic methods with gradients for neural model explanation.
problem Limited scalability of symbolic methods for large neural networks.
method Combines gradient-based methods with symbolic techniques using Integrated Gradients to focus on a subset of neurons.
result Produces sparser and higher saliency regions compared to gradient-based methods alone.
Post-hoc explanations improve CNNs by replacing final linear layer with k-means classifier.
problem CNNs lack accurate data representation in their built-in prototypes.
method Introduces k-means-based post-hoc explanations for CNNs, leveraging spatial consistency of convolutional receptive fields.
result Using shallower, less compressed feature activations improves semantic fidelity at the cost of slight predictive performance.
Researchers show how to manipulate Partial Dependence plots to deceive explanations of predictive models.
problem The robustness and trustworthiness of Partial Dependence (PD) explanations are compromised.
method Data poisoning using genetic and gradient algorithms to manipulate PD plots.
result PD explanations can be fooled and manipulated to mislead understanding of predictive models.
Local explanation methods for deep networks are found to be insensitive to parameter values.
problem Local explanation methods lack sensitivity to parameter values in deep neural networks.
method Investigated the sensitivity of local explanation methods (e.g., integrated gradients) to parameter values in randomly-initialized and learned networks.
result IG attributions for a random network and the actual network are uncorrelated when both factors (signs and baseline pixels) are accounted for.
Study shows group structures are crucial for financial model explanations.
problem Inconsistent explanations from existing explainable machine learning methods.
method Examined group structures in financial datasets and developed group versions of Shapley values.
result Group versions of Shapley values provide consistent explanations.
1-Lipschitz neural networks produce clearer, more focused Saliency Maps for explainable AI.
problem Noisy and limited Saliency Maps from traditional neural networks.
method Dual loss of optimal transport problem for 1-Lipschitz neural networks.
result Saliency Maps from 1-Lipschitz networks are highly concentrated and less noisy, aligning with human explanations.
ID-ExpO fine-tunes neural networks for more faithful explanations.
problem Improving the faithfulness of explanations for complex machine learning models.
method Differentiable insertion/deletion metric-aware regularizers for optimization.
result Fine-tuned predictors produce more faithful explanations.
Evaluates explanations of LTR models using decision paths and compares their accuracy.
problem Challenges in evaluating local explanations of LTR models due to lack of ground truth feature importance scores.
method Focuses on tree-based LTR models, extracts ground truth feature importance scores using decision paths, and compares them with explanation techniques.
result Explanation accuracy varies depending on the model and data point.
A fast method finds interpretable counterfactual explanations using class prototypes.
problem Finding understandable counterfactual explanations for classifier predictions.
method Using class prototypes, the method speeds up and improves interpretability of counterfactual instances.
result The method significantly speeds up and improves the interpretability of counterfactual explanations.
Extract class-specific subnetworks from neural models for better understanding and improved explanations.
problem Understanding and explaining the complex behavior of deep neural networks.
method For each semantic class, extract a class-specific subnetwork with a compressed structure that maintains comparable performance.
result Extracted subnetworks improve explanation saliency and adversarial example detection.
Dropout improves neural networks by accelerating gradient flow.
problem Understanding why dropout works and improving neural network performance.
method Proposed an optimization technique to push input towards saturation area of activation functions.
result Gradient acceleration in activation function (GAAF) improves image classification performance.
Modified BP attribution methods often ignore later layers' information, leading to misleading explanations.
problem Misleading explanations from modified BP methods ignoring later layers' information.
method Analysis of 9 modified BP methods including Deep Taylor Decomposition, LRP, Excitation BP, PatternAttribution, DeepLIFT, Deconv, RectGrad, Guided BP.
result Only DeepLIFT does not ignore later layers' information, providing a faithful explanation.
EXAGREE selects a stakeholder-aligned model to reduce conflicting explanations in machine learning.
problem Conflicting explanations from different attribution methods limit the adoption of machine learning models in safety-critical domains.
method EXAGREE is a two-stage framework that selects a Stakeholder-Aligned Explanation Model (SAEM) from a set of similar-performing models, maximizing Stakeholder-Machine Agreement (SMA).
result EXAGREE achieves simultaneous gains in faithfulness, plausibility, and fairness over baselines while preserving task accuracy.
Develops causal explanations for black-box classifiers.
problem Creating understandable explanations for complex models.
method Generative model with information-theoretic measures of causal influence.
result Learned latent factors influence classifier outputs.
GIG improves IG to explain diverse ML functions.
problem Explain diverse ML functions effectively.
method Generalized Integrated Gradients (GIG) method.
result GIG is the only correct method under reasonable axioms.
WassersteinGrad improves weather forecasting explanations by addressing geometric misalignment issues.
problem Improving explainability of autoregressive neural predictions on dynamic physical fields.
method WassersteinGrad, a geometric consensus method for averaged perturbed attribution maps.
result WassersteinGrad provides more accurate explanations for weather forecasting models.
New metric assesses reliability of AI explanations.
problem Unreliable AI explanations under realistic conditions.
method Explanation Reliability Index (ERI) metrics quantifying stability under four axioms.
result Widespread reliability failures in popular explanation methods.
New method explains neural network decisions and improves model performance.
problem Neural networks' opacity makes them hard to trust in critical applications.
method Efficiently explains and regularizes differentiable models by penalizing input gradients.
result Models generate faithful explanations and generalize better when conditions differ.
Unified framework for feature-based explanations using ANOVA and game theory.
problem Differences between feature-based explanations methods limit their applicability.
method Introduces a unified framework combining fANOVA and cooperative game theory.
result Uncovered similarities and differences between various explanation techniques.
RELAX provides first attribution-based explanations for representations.
problem Lack of methods to explain what influences learned representations.
method RELAX, a first approach for attribution-based explanations of representations, measuring similarities in representation space.
result Significantly outperforms gradient-based baseline and models uncertainty in explanations.
Survey of counterfactual explanations for time series classification.
problem Generating plausible and actionable counterfactuals for time series data.
method Review of various counterfactual generation methods for time series classification.
result Highlight unique challenges and strengths of existing methods.
CRITS improves time series classification with interpretable local explanations.
problem Lack of detailed explanations in time series classification models.
method CRITS uses convolutional kernels, max-pooling, and rectified linear units to extract feature weights.
result CRITS provides intrinsically interpretable local explanations without requiring gradients or random perturbations.
The paper shows averaging gradients leads to memorization, proposing an alternative algorithm to focus on invariances.
problem The principle that 'good explanations are hard to vary' in deep learning is investigated.
method Formalizing consistency for loss surface minima, proposing an alternative algorithm based on logical AND.
result The alternative algorithm prevents memorization and focuses on invariances.
Grad-CAM visualizes CNN predictions to explain model decisions.
problem Making CNN models transparent and understandable.
method Using class-specific gradient information to localize important regions.
result Improved understanding of CNN-based models, including image captioning and VQA.
ALMANACS benchmarks explainability methods on simulatability.
problem Evaluating the effectiveness of explainability methods for language models.
method ALMANACS is a simulatability benchmark that evaluates explainability methods on twelve safety-relevant topics.
result No explainability method outperforms the explanation-free control across all topics.
New findings show margins are not sufficient for explaining gradient boosting performance.
problem The inadequacy of margin explanations in explaining the performance of gradient boosting.
method Demonstrated and proved a stronger margin-based generalization bound for boosted classifiers.
result Proved a stronger margin-based generalization bound that explains the performance of modern gradient boosters.
Stacking improves deep neural network training efficiency.
problem Improving the efficiency of training deep neural networks.
method Proposes stacking as a form of accelerated gradient descent.
result Proves stacking provides accelerated training for certain deep linear residual networks.
Forward-Euler fails for simulating Wasserstein gradient flows with KL divergence.
problem Simulating Wasserstein gradient flows with forward-Euler discretization fails for KL divergence.
method Forward-Euler discretization for Wasserstein gradient flows with KL divergence.
result Forward-Euler discretization can be incorrect for Wasserstein gradient flows with KL divergence.
ViTaX provides formal guarantees for targeted explanations in safety-critical systems.
problem Need trustworthy explanations for safety-critical deep neural networks.
method Formal reachability analysis for targeted, semifactual explanations.
result First method to provide formally guaranteed explanations of model resilience.
Improved credit scoring model with explainability.
problem Making financial decisions based on loan applications.
method Extreme Gradient Boosting (XGBoost) model with 360-degree explanation framework.
result Model achieves state-of-the-art performance and provides understandable explanations.
Large batch sizes reduce gradient variance in DP-SGD, improving privacy.
problem Understanding why large batch sizes work in DP-SGD.
method Decomposed total gradient variance into subsampling and noise-induced variances, proving batch size independence in the limit.
result Large batch sizes reduce effective total gradient variance, improving privacy in DP-SGD.
New method makes neural network explanations more robust to attacks.
problem Vulnerability of feature attributions to adversarial attacks.
method Promotes Lipschitz continuity and smoothness in DNNs to improve robustness.
result Regularization and stochastic smoothing methods enhance attribution robustness.
TREX explains tree ensembles by identifying key training examples.
problem Identifying which training examples most influence tree ensemble predictions.
method TREX builds a surrogate model using a kernel that captures tree ensemble structure, approximating the original model.
result TREX provides accurate and effective explanations for tree ensembles.
Method generates visual explanations for similarity models without classification.
problem Lack of visual explanations for similarity models trained without classification loss.
method Gradient-based visual attention using learned feature embeddings.
result Attention maps improve model performance and can be used as constraints.
Deep nets' optima are surprisingly connected by simple paths.
problem Understanding the loss landscape of deep nets.
method Mathematical explanations based on generic properties of well-trained nets.
result Optima in deep nets are connected by piece-wise linear paths.
Improved sentiment analysis explanations using LRP for RNNs.
problem Creating understandable explanations for recurrent neural network predictions.
method Extending Layer-wise Relevance Propagation (LRP) to recurrent neural networks (RNNs), specifically to multiplicative connections in LSTMs and GRUs.
result Better explanation quality for sentiment analysis tasks using LRP compared to gradient-based methods.
The paper uses SHAP for interpreting machine learning models in hospital data.
problem Interpreting machine learning models in healthcare.
method SHAP for feature importance and feature packing techniques.
result SHAP provides better interpretability of machine learning models in healthcare.