Explainable AI: current status and future directions

I. INTRODUCTION

• Traditional AI is described as a black box that can answer yes-or-no questions without explaining how it obtained an answer.Source: §I · p. 1
• XAI is motivated by critical decisions where users need rationales and answers to “why,” “when,” or “where,” beyond the decision itself.Source: §I · pp. 1–2
• The survey groups XAI methods across text, image, audio, and video, noting that a method suited to one medium may not work for another.Source: §I · p. 2

A. OVERVIEW

• Black-box models in domains including socioeconomic justice, cyber forensics, and criminal justice are described as lacking trust and accountability.Source: A. OVERVIEW · p. 2

B. EXPLAINABILITY REQUIREMENT

• The paper’s refinery case links explanations to identifying sensors associated with unstable combustion and reducing investigation time after failures.Source: 1) Case Study 1: Oil Refinery Assets… · p. 2
• For video threat detection, the authors note that model skew or bias may lead to innocent visitors being detected as threats.Source: 2) Case Study 2: Video Threat Detection · p. 2

C. USAGE

• XAI is presented as a way to make expanding AI use more transparent and interpretable, including in medicine, autonomous driving, and criminal justice.The paper also notes ethical concerns around AI use in corporate and industrial settings.Source: C. USAGE · p. 3

D. KEY ISSUES WITH EXPLAINABLE ML

• Black-box models are difficult to understand when their functions are complex or proprietary; researchers also believe interpretability may reduce predictive accuracy.The paper links that belief to researchers developing deep-learning expertise without comparable XAI expertise.Source: D. KEY ISSUES WITH EXPLAINABLE ML · p. 3
• XAI justifications may not match a black-box model’s computation, even when both produce the same predictions.The paper notes that the models may use different features, and that explanations can include information irrelevant to the black box’s inferences.Source: D. KEY ISSUES WITH EXPLAINABLE ML · p. 3
• Saliency maps show where an image model focuses but not how image parts contribute to its prediction.The dog and musical-instrument explanations look essentially the same.Source: D. KEY ISSUES WITH EXPLAINABLE ML · p. 3, Figure 2 · p. 3
• Edge-focused saliency explanations can be identical across classes, even when the model is wrong, making a correct-class-only display misleading.The paper illustrates the concern with a Panda image tampered with adversarial noise and predicted as a Gibbon with high confidence [72].Source: D. KEY ISSUES WITH EXPLAINABLE ML · pp. 3–4, Figure 3 · p. 3

A. OBJECTIVES

• XAI aims to answer why and how an AI answer was obtained and when a system can fail, supporting trustworthiness, transparency, confidence, and informativeness.Source: A. OBJECTIVES · p. 4
• Transparency and informativeness require justifications that lay users can understand and models expressive enough to be human-understandable.Source: 1) Transparency and Informativeness · p. 4
• Transparency can help assess model performance and justifications, including vulnerabilities associated with false training.Source: 1) Transparency and Informativeness · p. 4
• Logical and scientific justifications are presented as reasons people may trust AI predictions and conclusions.Source: 2) Trust and confidence · p. 4
• The paper associates XAI with promoting fairness and helping mitigate bias in prediction interpretation.Source: 3) Bias Understanding and Fairness · p. 4

B. SCOPE

• XAI’s stated scopes include NLP, health care, engineering, and defense; the paper also names data protection, medical applications, defense, and banking.Banking examples include investigating fraudulent transactions and reducing false positives.Source: B. SCOPE · p. 4

IV. CLASSIFICATION TREE

• Transparent methods expose model decision processes, while post-hoc methods interpret nonlinear relationships or complex models after training.Post-hoc explanations may take the form of feature scores, rules, heat maps, or natural language.Source: §IV · p. 4, A. POSTHOC METHODS · p. 5

A. POSTHOC METHODS

• Post-hoc methods can be model-agnostic or model-specific, and local methods explain particular instances while global methods cover datasets.LIME perturbs samples, observes output changes, and fits a simple local approximation; model-specific methods depend on the learning algorithm or model structure.Source: A. POSTHOC METHODS · p. 5

B. TRANSPARENT METHODS

• Transparency is assessed through algorithmic transparency, decomposability, and simulatability, with complex inputs limiting interpretability.Human simulation also depends on model complexity; sparse matrices are easier to justify and visualize than dense matrices.Source: B. TRANSPARENT METHODS · p. 5
• Transparent-model examples include logistic regression, decision trees, KNN, rules, and Bayesian models, but complexity and audience knowledge constrain transparency.More decision-tree levels can improve algorithmic transparency while reducing simulability; ensembling can address poor generalization but makes trees less transparent. Complex Bayesian dependencies can also impair transparency and human simulation.Source: 1) Linear/Logistic Regression · p. 6, 2) Decision Trees · p. 6, 3) K-Nearest Neighbors · p. 6

C. MODEL SPECIFIC

• Feature importance identifies each feature’s impact on decisions, while feature correlations can inform medical diagnosis.The paper describes feature correlation in training data as one driving force for diagnosis.Source: 1) Feature Relevance · p. 6
• Condition-based explanations answer “why,” “why despite,” and “why given” questions with conditions, while “what if” supports counterfactual reasoning.A logical model converts user inputs into constraints and reports whether those constraints are satisfied.Source: 2) Condition based Explanation · p. 6
• Rules can translate insights into linguistic explanations for users and provide transparency across predictions, including those from complex neural networks.The paper describes rules as a way to present model insights to customers and non-expert users.Source: 2) Condition based Explanation · p. 6
• Saliency maps highlight image or video-frame regions significant to CNN decisions, using backpropagated gradients to set color intensity.The maps are presented as a way to expose aspects of deep neural network behavior.Source: 4) Feature based saliency map · p. 6

D. MODEL AGNOSTIC

• LIME provides local explanations for supervised models by sampling around an instance, weighting nearby examples, and selecting significant features.The paper describes a default of 5,000 normally distributed feature-vector samples and mentions lasso or PCA for feature selection.Source: 1) LIME- Local Interpretable… · p. 7
• Perturbation analyzes how feature changes affect target results and can be applied across AI/ML models, but becomes computationally expensive with many features.The paper also warns that perturbation can underestimate a feature’s contribution when that feature is saturated.Source: 2) Perturbation · p. 7
• Layer-wise Relevance Propagation explains complex neural networks by propagating predictions backward using designed rules..Source: 3) LRP: Layer-wise Relevance Propagation · p. 7
• Provenance and taxonomy induction use logical inference to justify results from partially derived results.The paper also identifies LRP, SHAP, provenance and taxonomy induction, and counterfactual explanations as applicable to text, image, audio, and video.Source: 4) Provenance and taxonomy induction · p. 7, D. MODEL AGNOSTIC · p. 7

V. XAI AND IMAGE

• Image-classification explanations such as LIME, SHAP, and LRP rank feature importance but do not determine optimal explanation size, account for feature dependence, or cover multiple prediction classes.The section distinguishes explanations of a model’s conclusion from those covering the prediction process, including training, and direct explanations from post-hoc analysis.Source: §V · p. 7

A. LIME

• LIME uses model predictions as labels to train a local explanation model, and sparse linear models can highlight weighted pixels for a particular class.The cited Inception example’s top predictions are Electric Guitar (0.32), Acoustic guitar (0.24), and Labrador (0.21) [73].Source: A. LIME · p. 7, Figure 7 · p. 7
• The paper describes its classification approach as self-explainable but says future work requires text-based presentation..Source: A. LIME · p. 8
• The section identifies validation of justification as necessary for greater reliability..Source: A. LIME · p. 8

B. SHAP (SHAPLEY ADDITIVE EXPLANATIONS)

• SHAP assigns feature-importance values to a particular prediction using Shapley values from coalitional game theory, while LIME and SHAP share several explanation limitations.The methods do not determine optimal explanation size, account for feature dependence, or explain more than one prediction class.Source: B. SHAP (SHAPLEY ADDITIVE EXPLANATIONS) · p. 8

C. COUNTERFACTUAL VISUAL EXPLANATIONS

• Counterfactuals explain why one outcome was selected over alternatives by identifying features whose presence or absence would change the model’s decision.The paper describes them as useful for examining how an interpretation was made, adjusting adverse decisions, and seeking intended predictions.Source: C. COUNTERFACTUAL VISUAL EXPLANATIONS · p. 8
• SEDC removes image segments until the prediction changes, whereas SEDC-T removes them until a predefined target class is reached.The paper illustrates minimal evidence with a warplane image and describes pertinent positives as minimally required factors and pertinent negatives as factors whose absence supports the decision.Source: C. COUNTERFACTUAL VISUAL EXPLANATIONS · p. 9, Figure 8 · p. 9

D. XAI AND HEALTHCARE

• Medical-image explanations include attribution methods such as heat maps, DeepTaylor, DeepExplain, and DeepLIFT, which assign or visualize feature contributions.DeepTaylor is described as generating positive evidence without changing the underlying CNN architecture; DeepLIFT decomposes a network prediction and assigns scores to input features.Source: 1) Explainability methods for… · p. 9
• Perturbation methods alter or mask input features and record effects on predictions; occlusion ranks image features, with tile size trading off against accuracy.The paper notes that full-pixel occlusion is computationally expensive and identifies 3 × 3 and 10 × 10 tiles as commonly useful.Source: 1) Explainability methods for… · pp. 9–10
• For brain imaging, CNN-based Alzheimer’s classification is described as robust, while post-hoc analysis can identify feature-related overfitting and support hyperparameter adjustment.Guided backpropagation, LRP, and DeepSHAP are identified for brain imaging and classification.Source: 2) XAI for Health care applications · p. 10
• The COVID-19 CT pipeline extracts lung information, classifies scans as positive or negative, localizes symptoms, and provides documented justification.The described model processes three consecutive 224 × 224 CT slices and generates six-channel segmentation for lung lobes and non-lung areas.Source: 2) XAI for Health care applications · p. 10

• NLP applications include sentiment analysis, hate-speech detection, and text summarization, using approaches such as decision trees, sequential modeling, logistic regression, bag-of-words, and skip-grams.The passage notes that word embeddings can improve black-box inference efficiency while making models less interpretable.Source: §VI · pp. 10–11
• Black-box NLP models raise concerns about interpretability and reliability, making explainability relevant to these applications.Source: §VI · p. 11
• The section identifies saliency heat maps, saliency highlights, declarative explanations, and natural-language inference as NLP explanation techniques.Source: §VI · p. 11

A. EXPLAINABILITY TECHNIQUES FOR NLP

• Feature-importance methods score individual features, including handcrafted, lexical, latent, and attention-based features; handcrafted features can be limited by local-optimum derivations.Source: A. EXPLAINABILITY TECHNIQUES FOR NLP · p. 11
• Example-driven explanations justify a prediction with supporting and opposing text examples, while provenance presents the reasoning steps leading to a result.The passage identifies provenance as particularly suited to automatic question-answering explanations.Source: A. EXPLAINABILITY TECHNIQUES FOR NLP · p. 11

B. VISUALIZATION TECHNIQUES FOR NLP

• NLP explanations can be presented as saliency maps, logic rules, trees, programs, or generated natural-language sentences.Saliency maps display scores for individual features, while natural-language explanations use domain-specific generative models.Source: B. VISUALIZATION TECHNIQUES FOR NLP · p. 11

C. XAI AND HATE SPEECH DETECTION

• For hate-speech detection, saliency and Shapley-based feature importance can expose words influencing classification, including a 94% confident misclassification of directed hate as offensive language.In that example, “bitch” and “faggot” contribute positively to the score but distract the classifier from the tweet’s hate.Source: C. XAI AND HATE SPEECH DETECTION · p. 11, Figure 14 · p. 12

VII. XAI AND VIDEO

• Video explanations can use frame-wise decomposition; LRP decomposes relevance pixel by pixel but requires access to the network’s internal architecture and weights.LIME instead approximates decisions using sampled inputs and can be applied to text, images, and tabular data.Source: §VII · p. 12

VIII. XAI AND AUDIO

• For audio, NLP-based XAI can follow conversion of speech waves to text, while a user study found virtual-agent visualizations gained more trust than text- or voice-only interaction.The study used a speech-recognition model classifying audio keywords from spectrograms; LIME is also applied to visualize voice classification.Source: §VIII · pp. 12–13

IX. XAI AND MULTIMODAL DATA

• Sensor degradation readings can support aviation failure diagnosis and remaining-useful-life prediction, while explanations justify diagnoses and are said to improve reliability and save cost.The discussion also describes failure diagnosis as making ML models more transparent and interpretable.Source: §IX · p. 13
• Predictive maintenance combines data collection, cleaning, feature selection, diagnosis, and explanation with validation, offering maintenance teams insights about component disposal.Figure 16 depicts an automated failure-diagnosis model.Source: §IX · p. 13, Figure 16 · p. 14

X. CONCLUSION

• The conclusion covers XAI objectives including transparency, fairness, bias, and confidence, and its scope across NLP, medical, defense, and engineering domains.It emphasizes XAI as important for real-time AI/ML applications.Source: §X · p. 13
• The survey discusses transparent and post hoc explainability methods and applies XAI approaches to image, text, video, audio, and multimodal data.It presents conceptual explanations and examples of the methodologies.Source: §X · p. 13
• The survey frames XAI as a tool for responsible AI, associated with trust, confidence, fairness, and transparency.It distinguishes transparent ML models from post hoc methods that make models more interpretable.Source: §X · p. 13