Content deleted Content added
No edit summary |
LIME acronym |
||
| (459 intermediate revisions by more than 100 users not shown) | |||
Line 1:
{{short description|AI whose outputs can be understood by humans}}
{{other uses of|XAI}}
{{artificial intelligence}}
Within [[artificial intelligence]] (AI), '''explainable AI''' ('''XAI'''), often overlapping with '''interpretable AI''' or '''explainable machine learning''' ('''XML'''), is a field of research that explores methods that provide humans with the ability of intellectual oversight over AI algorithms.<ref>{{Cite journal|last=Longo|first=Luca|display-authors=etal|date=2024 |title=Explainable Artificial Intelligence (XAI) 2.0: A manifesto of open challenges and interdisciplinary research directions|url=https://www.sciencedirect.com/science/article/pii/S1566253524000794 |journal=Information Fusion|volume=106|article-number=102301 |doi=10.1016/j.inffus.2024.102301|arxiv=2310.19775}}</ref><ref>{{Cite journal |last=Mihály |first=Héder |date=2023 |title=Explainable AI: A Brief History of the Concept |url=https://ercim-news.ercim.eu/images/stories/EN134/EN134-web.pdf |journal=ERCIM News |issue=134 |pages=9–10}}</ref> The main focus is on the [[automated reasoning|reasoning]] behind the decisions or predictions made by the AI algorithms,<ref>{{Cite journal |last1=Phillips |first1=P. Jonathon |last2=Hahn |first2=Carina A. |last3=Fontana |first3=Peter C. |last4=Yates |first4=Amy N. |last5=Greene |first5=Kristen |last6=Broniatowski |first6=David A. |last7=Przybocki |first7=Mark A. |date=2021-09-29 |title=Four Principles of Explainable Artificial Intelligence |url=https://doi.org/10.6028/NIST.IR.8312 |journal=NIST |doi=10.6028/nist.ir.8312}}</ref> to make them more understandable and transparent.<ref>{{Cite journal|last1=Vilone|first1=Giulia|last2=Longo|first2=Luca|title=Notions of explainability and evaluation approaches for explainable artificial intelligence|url=https://www.sciencedirect.com/science/article/pii/S1566253521001093|journal=Information Fusion|year=2021|volume= December 2021 - Volume 76 |pages=89–106|doi=10.1016/j.inffus.2021.05.009}}</ref> This addresses users' requirement to assess safety and scrutinize the automated decision making in applications.<ref name="auto">{{Cite journal |last1=Confalonieri |first1=Roberto |last2=Coba |first2=Ludovik |last3=Wagner |first3=Benedikt |last4=Besold |first4=Tarek R. |date=January 2021 |title=A historical perspective of explainable Artificial Intelligence |url=https://wires.onlinelibrary.wiley.com/doi/10.1002/widm.1391 |journal=WIREs Data Mining and Knowledge Discovery |language=en |volume=11 |issue=1 |article-number=e1391 |doi=10.1002/widm.1391 |issn=1942-4787|hdl=11577/3471605 |hdl-access=free }}</ref> XAI counters the "[[black box]]" tendency of machine learning, where even the AI's designers cannot explain why it arrived at a specific decision.<ref>{{Cite journal |last=Castelvecchi |first=Davide |date=2016-10-06 |title=Can we open the black box of AI? |journal=Nature |language=en |volume=538 |issue=7623 |pages=20–23 |doi=10.1038/538020a |pmid=27708329 |bibcode=2016Natur.538...20C |s2cid=4465871 |issn=0028-0836|doi-access=free }}</ref><ref name=guardian>{{cite news|last1=Sample|first1=Ian|title=Computer says no: why making AIs fair, accountable and transparent is crucial|url=https://www.theguardian.com/science/2017/nov/05/computer-says-no-why-making-ais-fair-accountable-and-transparent-is-crucial|access-date=30 January 2018|work=The Guardian |date=5 November 2017|language=en}}</ref>
XAI hopes to help users of AI-powered systems perform more effectively by improving their understanding of how those systems reason.<ref>{{Cite journal|last=Alizadeh|first=Fatemeh|date=2021|title=I Don't Know, Is AI Also Used in Airbags?: An Empirical Study of Folk Concepts and People's Expectations of Current and Future Artificial Intelligence|url=https://www.researchgate.net/publication/352638184|journal=Icom|volume=20 |issue=1 |pages=3–17 |doi=10.1515/icom-2021-0009|s2cid=233328352}}</ref> XAI may be an implementation of the social [[right to explanation]].<ref name=":0">{{Cite journal|last1=Edwards|first1=Lilian|last2=Veale|first2=Michael|date=2017|title=Slave to the Algorithm? Why a 'Right to an Explanation' Is Probably Not the Remedy You Are Looking For|journal=Duke Law and Technology Review|volume=16|pages=18|ssrn=2972855}}</ref> Even if there is no such legal right or regulatory requirement, XAI can improve the [[user experience]] of a product or service by helping end users trust that the AI is making good decisions.<ref>{{Cite web |last=Do Couto |first=Mark |date=February 22, 2024 |title=Entering the Age of Explainable AI |url=https://tdwi.org/Articles/2024/02/22/ADV-ALL-Entering-the-Age-of-Explainable-AI.aspx |access-date=2024-09-11 |website=TDWI}}</ref> XAI aims to explain what has been done, what is being done, and what will be done next, and to unveil which information these actions are based on.<ref name=":3">{{Cite journal|last1=Gunning|first1=D.|last2=Stefik|first2=M.|last3=Choi|first3=J.|last4=Miller|first4=T.|last5=Stumpf|first5=S.|last6=Yang|first6=G.-Z.|date=2019-12-18|title=XAI-Explainable artificial intelligence|url=https://openaccess.city.ac.uk/id/eprint/23405/|journal=Science Robotics|language=en|volume=4|issue=37|pages=eaay7120|doi=10.1126/scirobotics.aay7120|pmid=33137719|issn=2470-9476|doi-access=free}}</ref> This makes it possible to confirm existing knowledge, challenge existing knowledge, and generate new assumptions.<ref>{{Cite journal|last1=Rieg|first1=Thilo|last2=Frick|first2=Janek|last3=Baumgartl|first3=Hermann|last4=Buettner|first4=Ricardo|date=2020-12-17|title=Demonstration of the potential of white-box machine learning approaches to gain insights from cardiovascular disease electrocardiograms|journal=PLOS ONE|language=en|volume=15|issue=12|pages=e0243615|doi=10.1371/journal.pone.0243615|issn=1932-6203|pmc=7746264|pmid=33332440|bibcode=2020PLoSO..1543615R|doi-access=free}}</ref>
==Background==
[[Machine learning]] (ML) algorithms used in AI can be categorized as [[White-box testing|white-box]] or [[Black box|black-box]].<ref>{{Cite journal|last1=Vilone|first1=Giulia|last2=Longo|first2=Luca|title= Classification of Explainable Artificial Intelligence Methods through Their Output Formats |journal=Machine Learning and Knowledge Extraction|year=2021|volume=3|issue=3|pages=615–661|doi=10.3390/make3030032|doi-access=free }}</ref> White-box models provide results that are understandable to experts in the ___domain. Black-box models, on the other hand, are extremely hard to explain and may not be understood even by ___domain experts.<ref>{{Cite journal|last=Loyola-González|first=O.|date=2019|title=Black-Box vs. White-Box: Understanding Their Advantages and Weaknesses From a Practical Point of View|journal=IEEE Access|volume=7|pages=154096–154113|doi=10.1109/ACCESS.2019.2949286|bibcode=2019IEEEA...7o4096L |issn=2169-3536|doi-access=free}}</ref> XAI algorithms follow the three principles of transparency, interpretability, and explainability.
* A model is transparent "if the processes that extract model parameters from training data and generate labels from testing data can be described and motivated by the approach designer."<ref name=":4">{{Cite journal|last1=Roscher|first1=R.|last2=Bohn|first2=B.|last3=Duarte|first3=M. F.|last4=Garcke|first4=J.|date=2020|title=Explainable Machine Learning for Scientific Insights and Discoveries|journal=IEEE Access|volume=8|pages=42200–42216|doi=10.1109/ACCESS.2020.2976199|arxiv=1905.08883 |bibcode=2020IEEEA...842200R |issn=2169-3536|doi-access=free}}</ref>
* Interpretability describes the possibility of comprehending the ML model and presenting the underlying basis for decision-making in a way that is understandable to humans.<ref name="Interpretable machine learning: def">{{cite journal|last1=Murdoch|first1=W. James|last2=Singh|first2=Chandan|last3=Kumbier|first3=Karl|last4=Abbasi-Asl|first4=Reza|last5=Yu|first5=Bin|date=2019-01-14|title=Interpretable machine learning: definitions, methods, and applications|journal=Proceedings of the National Academy of Sciences of the United States of America|volume=116|issue=44|pages=22071–22080|arxiv=1901.04592|doi=10.1073/pnas.1900654116|pmid=31619572|pmc=6825274|bibcode= |doi-access=free}}</ref><ref name="Lipton 31–57">{{Cite journal|last=Lipton|first=Zachary C.|date=June 2018|title=The Mythos of Model Interpretability: In machine learning, the concept of interpretability is both important and slippery.|journal=Queue|language=en|volume=16|issue=3|pages=31–57|doi=10.1145/3236386.3241340|issn=1542-7730|doi-access=free|arxiv=1606.03490}}</ref><ref>{{Cite web|date=2019-10-22|title=Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI|url=https://deepai.org/publication/explainable-artificial-intelligence-xai-concepts-taxonomies-opportunities-and-challenges-toward-responsible-ai|access-date=2021-01-13|website=DeepAI}}</ref>
* Explainability is a concept that is recognized as important, but a consensus definition is not yet available;<ref name=":4" /> one possibility is "the collection of features of the interpretable ___domain that have contributed, for a given example, to producing a decision (e.g., classification or regression)".<ref>{{Cite journal|date=2018-02-01|title=Methods for interpreting and understanding deep neural networks|journal=Digital Signal Processing|language=en|volume=73|pages=1–15|doi=10.1016/j.dsp.2017.10.011|issn=1051-2004|doi-access=free|last1=Montavon|first1=Grégoire|last2=Samek|first2=Wojciech|last3=Müller|first3=Klaus-Robert|arxiv=1706.07979 |bibcode=2018DSP....73....1M |author-link3=Klaus-Robert Müller}}</ref>
In summary, Interpretability refers to the user's ability to understand model outputs, while Model Transparency includes Simulatability (reproducibility of predictions), Decomposability (intuitive explanations for parameters), and Algorithmic Transparency (explaining how algorithms work). Model Functionality focuses on textual descriptions, visualization, and local explanations, which clarify specific outputs or instances rather than entire models. All these concepts aim to enhance the comprehensibility and usability of AI systems.<ref name="NCB23">{{Cite web|author = Notovich, Aviv., Chalutz-Ben Gal, Hila, & Ben-Gal, Irad (2023). |title = Explainable Artificial Intelligence (XAI): Motivation, Terminology, and Taxonomy.|url = https://www.iradbengal.sites.tau.ac.il/_files/ugd/901879_fcdc9f3f44dc41968ba476a66f60f078.pdf|publisher = In Machine Learning for Data Science Handbook: Data Mining and Knowledge Discovery Handbook (pp. 971-985). Cham: Springer International Publishing.}}</ref>
If algorithms fulfill these principles, they provide a basis for justifying decisions, tracking them and thereby verifying them, improving the algorithms, and exploring new facts.<ref>{{Cite journal|last1=Adadi|first1=A.|last2=Berrada|first2=M.|date=2018|title=Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI)|journal=IEEE Access|volume=6|pages=52138–52160|doi=10.1109/ACCESS.2018.2870052|bibcode=2018IEEEA...652138A |issn=2169-3536|doi-access=free}}</ref>
Sometimes it is also possible to achieve a high-accuracy result with white-box ML algorithms. These algorithms have an interpretable structure that can be used to explain predictions.<ref name=":6">{{Cite journal|last=Rudin|first=Cynthia|date=2019|title=Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead|journal=Nature Machine Intelligence|language=en|volume=1|issue=5|pages=206–215|doi=10.1038/s42256-019-0048-x|pmid=35603010 |pmc=9122117 |arxiv=1811.10154|issn=2522-5839|doi-access=free}}</ref> Concept Bottleneck Models, which use concept-level abstractions to explain model reasoning, are examples of this and can be applied in both image<ref name="Koh Nguyen Tang Mussmann Pierson Kim Liang 2020">{{Cite conference|last1=Koh|first1=P. W.|last2=Nguyen|first2=T.|last3=Tang|first3=Y. S.|last4=Mussmann|first4=S.|last5=Pierson|first5=E.|last6=Kim|first6=B.|last7=Liang|first7=P.|date=November 2020|title=Concept bottleneck models|book-title=International Conference on Machine Learning|pages=5338–5348|publisher=PMLR}}</ref> and text<ref name="Ludan Lyu Yang Dugan Yatskar Callison-Burch 2023">{{Cite arXiv|last1=Ludan|first1=J. M.|last2=Lyu|first2=Q.|last3=Yang|first3=Y.|last4=Dugan|first4=L.|last5=Yatskar|first5=M.|last6=Callison-Burch|first6=C.|date=2023|title=Interpretable-by-Design Text Classification with Iteratively Generated Concept Bottleneck|class=cs.CL |eprint=2310.19660}}</ref> prediction tasks. This is especially important in domains like [[medicine]], [[Defense industry|defense]], [[finance]], and [[law]], where it is crucial to understand decisions and build trust in the algorithms.<ref name=":3" /> Many researchers argue that, at least for [[supervised machine learning]], the way forward is [[symbolic regression]], where the algorithm searches the space of mathematical expressions to find the model that best fits a given dataset.<ref name="Wenninger Kaymakci Wiethe 2022 p=118300">{{cite journal | last1=Wenninger | first1=Simon | last2=Kaymakci | first2=Can | last3=Wiethe | first3=Christian | title=Explainable long-term building energy consumption prediction using QLattice | journal=Applied Energy | publisher=Elsevier BV | volume=308 | year=2022 | issn=0306-2619 | doi=10.1016/j.apenergy.2021.118300 | article-number=118300| bibcode=2022ApEn..30818300W | s2cid=245428233 }}</ref><ref name="Christiansen Wilstrup Hedley 2022 p.">{{cite journal | last1=Christiansen | first1=Michael | last2=Wilstrup | first2=Casper | last3=Hedley | first3=Paula L. | title=Explainable "white-box" machine learning is the way forward in preeclampsia screening | journal=American Journal of Obstetrics and Gynecology | publisher=Elsevier BV | year=2022 | volume=227 | issue=5 | issn=0002-9378 | doi=10.1016/j.ajog.2022.06.057 | page=791| pmid=35779588 | s2cid=250160871 }}</ref><ref name="Wilstup Cave p.">{{citation | last1=Wilstup | first1=Casper | last2=Cave | first2=Chris | title=Combining symbolic regression with the Cox proportional hazards model improves prediction of heart failure deaths | publisher=Cold Spring Harbor Laboratory | date=2021-01-15 | doi=10.1101/2021.01.15.21249874 | page=| s2cid=231609904 | doi-access=free }}</ref>
AI systems optimize behavior to satisfy a mathematically specified goal system chosen by the system designers, such as the command "maximize the accuracy of [[sentiment analysis|assessing how positive]] film reviews are in the test dataset." The AI may learn useful general rules from the test set, such as "reviews containing the word "horrible" are likely to be negative." However, it may also learn inappropriate rules, such as "reviews containing '[[Daniel Day-Lewis]]' are usually positive"; such rules may be undesirable if they are likely to fail to generalize outside the training set, or if people consider the rule to be "cheating" or "unfair." A human can audit rules in an XAI to get an idea of how likely the system is to generalize to future real-world data outside the test set.<ref name="science">{{cite journal|date=5 July 2017|title=How AI detectives are cracking open the black box of deep learning|url=https://www.science.org/content/article/how-ai-detectives-are-cracking-open-black-box-deep-learning|journal=Science|language=en|access-date=30 January 2018}}.</ref>
==
Cooperation between [[Agency (sociology)|agents]] – in this case, [[Algorithm|algorithms]] and humans – depends on trust. If humans are to accept algorithmic prescriptions, they need to trust them. Incompleteness in formal trust criteria is a barrier to optimization. Transparency, interpretability, and explainability are intermediate goals on the road to these more comprehensive trust criteria.<ref name=dosilovic2018>{{cite conference |url=http://docs.mipro-proceedings.com/dsdc/dsdc_11_4754.pdf |title=Explainable Artificial Intelligence: A Survey |last1=Dosilovic |first1=Filip |last2=Brcic |first2=Mario |last3=Hlupic |first3=Nikica |date=2018-05-25 |book-title=MIPRO 2018 - 41st International Convention Proceedings |pages=210–215 |___location=Opatija, Croatia |conference=MIPRO 2018 |doi=10.23919/MIPRO.2018.8400040 |isbn=978-953-233-095-3 |access-date=2018-12-09 |archive-date=2018-12-10 |archive-url=https://web.archive.org/web/20181210110820/http://docs.mipro-proceedings.com/dsdc/dsdc_11_4754.pdf |url-status=dead }}</ref> This is particularly relevant in medicine,<ref>{{Cite journal |last1=Bernal |first1=Jose |last2=Mazo |first2=Claudia |date=2022-10-11 |title=Transparency of Artificial Intelligence in Healthcare: Insights from Professionals in Computing and Healthcare Worldwide |journal=Applied Sciences |language=en |volume=12 |issue=20 |pages=10228 |doi=10.3390/app122010228 |issn=2076-3417|doi-access=free }}</ref> especially with [[clinical decision support system]]s (CDSS), in which medical professionals should be able to understand how and why a machine-based decision was made in order to trust the decision and augment their decision-making process.<ref>{{Cite journal |last1=Antoniadi |first1=Anna Markella |last2=Du |first2=Yuhan |last3=Guendouz |first3=Yasmine |last4=Wei |first4=Lan |last5=Mazo |first5=Claudia |last6=Becker |first6=Brett A. |last7=Mooney |first7=Catherine |date=January 2021 |title=Current Challenges and Future Opportunities for XAI in Machine Learning-Based Clinical Decision Support Systems: A Systematic Review |journal=Applied Sciences |language=en |volume=11 |issue=11 |pages=5088 |doi=10.3390/app11115088 |issn=2076-3417|doi-access=free }}</ref>
AI systems sometimes learn undesirable tricks that do an optimal job of satisfying explicit pre-programmed goals on the training data but do not reflect the more nuanced implicit desires of the human system designers or the full complexity of the ___domain data. For example, a 2017 system tasked with [[image recognition]] learned to "cheat" by looking for a copyright tag that happened to be associated with horse pictures rather than learning how to tell if a horse was actually pictured.<ref name=guardian/> In another 2017 system, a [[supervised learning]] AI tasked with grasping items in a virtual world learned to cheat by placing its manipulator between the object and the viewer in a way such that it falsely appeared to be grasping the object.<ref>{{cite news|title=DeepMind Has Simple Tests That Might Prevent Elon Musk's AI Apocalypse|url=https://www.bloomberg.com/news/articles/2017-12-11/deepmind-has-simple-tests-that-might-prevent-elon-musk-s-ai-apocalypse|access-date=30 January 2018|work=Bloomberg.com|date=11 December 2017|language=en}}</ref><ref>{{cite news|title=Learning from Human Preferences|url=https://blog.openai.com/deep-reinforcement-learning-from-human-preferences/|access-date=30 January 2018|work=OpenAI Blog|date=13 June 2017}}</ref>
One transparency project, the [[DARPA]] XAI program, aims to produce "[[glass box]]" models that are explainable to a "[[human-in-the-loop]]" without greatly sacrificing AI performance. Human users of such a system can understand the AI's cognition (both in real-time and after the fact) and can determine whether to trust the AI.<ref>{{cite web|title=Explainable Artificial Intelligence (XAI)|url=https://www.darpa.mil/program/explainable-artificial-intelligence|website=DARPA|access-date=17 July 2017}}</ref> Other applications of XAI are [[knowledge extraction]] from black-box models and model comparisons.<ref>{{cite journal|last=Biecek|first=Przemyslaw|title= DALEX: explainers for complex predictive models|journal=Journal of Machine Learning Research|volume=19|pages=1–5|arxiv=1806.08915|date=23 June 2018}}</ref> In the context of monitoring systems for ethical and socio-legal compliance, the term "glass box" is commonly used to refer to tools that track the inputs and outputs of the system in question, and provide value-based explanations for their behavior. These tools aim to ensure that the system operates in accordance with ethical and legal standards, and that its decision-making processes are transparent and accountable. The term "glass box" is often used in contrast to "black box" systems, which lack transparency and can be more difficult to monitor and regulate.<ref>Rai, Arun. "Explainable AI: From black box to glass box." Journal of the Academy of Marketing Science 48 (2020): 137-141.</ref>
The term is also used to name a voice assistant that produces counterfactual statements as explanations.<ref name="SokolFlach2018">{{cite book|last1=Sokol|first1=Kacper|last2=Flach|first2=Peter|title=Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence|chapter=Glass-Box: Explaining AI Decisions With Counterfactual Statements Through Conversation With a Voice-enabled Virtual Assistant|year=2018|pages=5868–5870|doi=10.24963/ijcai.2018/865|isbn=9780999241127|s2cid=51608978}}</ref>
== Explainability and interpretability techniques ==
There is a subtle difference between the terms explainability and interpretability in the context of AI.<ref>{{cite journal |last1=Broniatowski |first1=David A. |title=Psychological Foundations of Explainability and Interpretability in Artificial Intelligence |journal=NIST Pubs |date=2021 |url=https://www.nist.gov/publications/psychological-foundations-explainability-and-interpretability-artificial-intelligence}}</ref>
{| class="wikitable sortable"
|-
! Term !! Definition !! Source
|-
| Interpretability || ''"level of understanding how the underlying (AI) technology works"'' || ISO/IEC TR 29119-11:2020(en), 3.1.42<ref name=ISO29119>{{cite ISO standard |csnumber=79016 |title=ISO/IEC TR 29119-11:2020, Software and systems engineering, Software testing, Part 11: Guidelines on the testing of AI-based systems |date=2020 |publisher=[[International Organization for Standardization|ISO]] |access-date=25 November 2023}}</ref>
|-
| Explainability || ''"level of understanding how the AI-based system ... came up with a given result"'' || ISO/IEC TR 29119-11:2020(en), 3.1.31<ref name=ISO29119/>
|}
Some explainability techniques don't involve understanding how the model works, and may work across various AI systems. Treating the model as a black box and analyzing how marginal changes to the inputs affect the result sometimes provides a sufficient explanation.
=== Explainability ===
Explainability is useful for ensuring that AI models are not making decisions based on irrelevant or otherwise unfair criteria. For [[Statistical classification|classification]] and [[Regression analysis|regression]] models, several popular techniques exist:
* ''Partial dependency plots'' show the marginal effect of an input feature on the predicted outcome.
* ''SHAP'' (SHapley Additive exPlanations) enables visualization of the contribution of each input feature to the output. It works by calculating [[Shapley value|Shapley values]], which measure the average marginal contribution of a feature across all possible combinations of features.<ref>{{Cite web |last=Verma |first=Yugesh |date=2021-12-25 |title=Complete Guide to SHAP - SHAPley Additive exPlanations for Practitioners |url=https://analyticsindiamag.com/a-complete-guide-to-shap-shapley-additive-explanations-for-practitioners/ |access-date=2024-07-10 |website=Analytics India Magazine |language=en-US}}</ref>
* ''Feature importance'' estimates how important a feature is for the model. It is usually done using ''permutation importance'', which measures the performance decrease when it the feature value randomly shuffled across all samples.
* ''LIME'' (Local Interpretable Model-Agnostic Explanations method) approximates locally a model's outputs with a simpler, interpretable model.<ref>{{Cite web |last=Rothman |first=Denis |date=2020-10-07 |title=Exploring LIME Explanations and the Mathematics Behind It |url=https://www.codemotion.com/magazine/ai-ml/lime-explainable-ai/ |access-date=2024-07-10 |website=Codemotion Magazine |language=en-US}}</ref>
* ''[[Multitask learning]]'' provides a large number of outputs in addition to the target classification. These other outputs can help developers deduce what the network has learned.<ref>{{Cite book |last=Christian |first=Brian |title=[[The Alignment Problem]]: Machine learning and human values |date=2020 |publisher=W. W. Norton & Company |isbn=978-0-393-86833-3 |chapter=TELL ME EVERYTHING: MULTITASK NETS}}</ref>
For images, [[Saliency map|saliency maps]] highlight the parts of an image that most influenced the result.<ref>{{Cite web |last=Sharma |first=Abhishek |date=2018-07-11 |title=What Are Saliency Maps In Deep Learning? |url=https://analyticsindiamag.com/what-are-saliency-maps-in-deep-learning/ |access-date=2024-07-10 |website=Analytics India Magazine |language=en-US}}</ref>
Systems that are expert or knowledge based are software systems that are made by experts. This system consists of a knowledge based encoding for the ___domain knowledge. This system is usually modeled as production rules, and someone uses this knowledge base which the user can question the system for knowledge. In expert systems, the language and explanations are understood with an explanation for the reasoning or a problem solving activity.<ref name="auto"/>
However, these techniques are not very suitable for [[Language model|language models]] like [[Generative pre-trained transformer|generative pretrained transformers]]. Since these models generate language, they can provide an explanation, but which may not be reliable. Other techniques include [[Attention (machine learning)|attention]] analysis (examining how the model focuses on different parts of the input), probing methods (testing what information is captured in the model's representations), causal tracing (tracing the flow of information through the model) and circuit discovery (identifying specific subnetworks responsible for certain behaviors). Explainability research in this area overlaps significantly with interpretability and [[AI alignment|alignment]] research.<ref>{{cite arXiv |last1=Luo |first1=Haoyan |title=From Understanding to Utilization: A Survey on Explainability for Large Language Models |date=2024-02-21 |eprint=2401.12874 |last2=Specia |first2=Lucia|class=cs.CL }}</ref>
=== Interpretability ===
[[File:Grokking modular addition.jpg|thumb|upright=1.2|[[Grokking (machine learning)|Grokking]] is an example of phenomenon studied in interpretability. It involves a model that initially memorizes all the answers ([[overfitting]]), but later adopts an algorithm that generalizes to unseen data.<ref>{{Cite web |last=Ananthaswamy |first=Anil |date=2024-04-12 |title=How Do Machines 'Grok' Data? |url=https://www.quantamagazine.org/how-do-machines-grok-data-20240412/ |access-date=2025-01-21 |website=Quanta Magazine |language=en}}</ref>]]
Scholars sometimes use the term "[[mechanistic interpretability]]" to refer to the process of [[Reverse engineering|reverse-engineering]] [[artificial neural networks]] to understand their internal decision-making mechanisms and components, similar to how one might analyze a complex machine or computer program.<ref>{{Cite web |last=Olah |first=Chris |date=June 27, 2022 |title=Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases |url=https://www.transformer-circuits.pub/2022/mech-interp-essay |access-date=2024-07-10 |website=www.transformer-circuits.pub}}</ref>
Interpretability research often focuses on generative pretrained transformers. It is particularly relevant for [[AI safety]] and [[AI alignment|alignment]], as it may enable to identify signs of undesired behaviors such as [[sycophancy]], deceptiveness or bias, and to better steer AI models.<ref>{{Cite web |last=Mittal |first=Aayush |date=2024-06-17 |title=Understanding Sparse Autoencoders, GPT-4 & Claude 3 : An In-Depth Technical Exploration |url=https://www.unite.ai/understanding-sparse-autoencoders-gpt-4-claude-3-an-in-depth-technical-exploration/ |access-date=2024-07-10 |website=Unite.AI |language=en-US}}</ref>
Studying the interpretability of the most advanced [[Foundation model|foundation models]] often involves searching for an automated way to identify "features" in generative pretrained transformers. In a [[Neural network (machine learning)|neural network]], a feature is a pattern of neuron activations that corresponds to a concept. A compute-intensive technique called "[[dictionary learning]]" makes it possible to identify features to some degree. Enhancing the ability to identify and edit features is expected to significantly improve the [[AI safety|safety]] of [[Frontier model|frontier AI models]].<ref>{{Cite web |last=Ropek |first=Lucas |date=2024-05-21 |title=New Anthropic Research Sheds Light on AI's 'Black Box' |url=https://gizmodo.com/new-anthropic-research-sheds-light-on-ais-black-box-1851491333 |access-date=2024-05-23 |website=Gizmodo |language=en}}</ref><ref>{{Cite magazine |last=Perrigo |first=Billy |date=2024-05-21 |title=Artificial Intelligence Is a 'Black Box.' Maybe Not For Long |url=https://time.com/6980210/anthropic-interpretability-ai-safety-research/ |access-date=2024-05-24 |magazine=Time |language=en}}</ref>
For [[Convolutional neural network|convolutional neural networks]], [[DeepDream]] can generate images that strongly activate a particular neuron, providing a visual hint about what the neuron is trained to identify.<ref>{{Cite magazine |last=Barber |first=Gregory |title=Inside the 'Black Box' of a Neural Network |url=https://www.wired.com/story/inside-black-box-of-neural-network/ |access-date=2024-07-10 |magazine=Wired |language=en-US |issn=1059-1028}}</ref>
== History and methods ==
During the 1970s to 1990s, [[Symbolic artificial intelligence|symbolic reasoning systems]], such as [[Mycin|MYCIN]],<ref>{{cite journal |last1=Fagan |first1=L. M. |last2=Shortliffe |first2=E. H. |last3=Buchanan |first3=B. G. |year=1980 |title=Computer-based medical decision making: from MYCIN to VM |journal=Automedica |volume=3 |issue=2 |pages=97–108}}</ref> GUIDON,<ref>{{Cite book| publisher = The MIT Press| last = Clancey| first = William| title = Knowledge-Based Tutoring: The GUIDON Program| ___location = Cambridge, Massachusetts| date = 1987}}</ref> SOPHIE,<ref>{{Cite book| publisher = Academic Press| isbn = 0-12-648680-8| last1 = Brown| first1 = John S.| last2 = Burton| first2 = R. R.| last3 = De Kleer| first3 = Johan| title = Intelligent Tutoring Systems| chapter = Pedagogical, natural language, and knowledge engineering techniques in SOPHIE I, II, and III| date = 1982}}</ref> and PROTOS<ref>{{Cite book| publisher = Morgan Kaufmann Publishers Inc.| isbn = 1-55860-119-8| volume = 3| pages = 112–139| last1 = Bareiss| first1 = Ray| last2 = Porter| first2 = Bruce| last3 = Weir| first3 = Craig| last4 = Holte| first4 = Robert| title = Machine Learning| chapter = Protos: An Exemplar-Based Learning Apprentice| date = 1990}}</ref><ref name=":1">{{Cite book |last=Bareiss |first=Ray |title=Exemplar-Based Knowledge Acquisition: A Unified Approach to Concept Representation, Classification, and Learning |publisher=[[Academic Press]] |year=1989 |series=Perspectives in Artificial Intelligence}}</ref> could represent, reason about, and explain their reasoning for diagnostic, instructional, or machine-learning (explanation-based learning) purposes. MYCIN, developed in the early 1970s as a research prototype for diagnosing [[bacteremia]] infections of the bloodstream, could explain<ref>{{cite book |last1=Van Lent |first1=M. |last2=Fisher |first2=W. |last3=Mancuso |first3=M. |date=July 2004 |chapter=An explainable artificial intelligence system for small-unit tactical behavior |title=Proceedings of the National Conference on Artificial Intelligence |pages=900–907 |___location=San Jose, CA |publisher=AAAI Press |isbn=0262511835}}</ref> which of its hand-coded rules contributed to a diagnosis in a specific case. Research in [[intelligent tutoring systems]] resulted in developing systems such as SOPHIE that could act as an "articulate expert", explaining problem-solving strategy at a level the student could understand, so they would know what action to take next. For instance, SOPHIE could explain the qualitative reasoning behind its electronics troubleshooting, even though it ultimately relied on the [[SPICE]] circuit simulator. Similarly, GUIDON added tutorial rules to supplement MYCIN's ___domain-level rules so it could explain the strategy for medical diagnosis. Symbolic approaches to machine learning relying on explanation-based learning, such as PROTOS, made use of explicit representations of explanations expressed in a dedicated explanation language, both to explain their actions and to acquire new knowledge.<ref name=":1" />
In the 1980s through the early 1990s, [[Truth-maintenance systems|truth maintenance systems]] (TMS) extended the capabilities of causal-reasoning, [[Rule-based system|rule-based]], and logic-based inference systems.<ref>{{Cite book| edition = Second| publisher = Prentice Hall, Pearson Education| isbn = 0-13-790395-2| last1 = Russell| first1 = Stuart| last2 = Norvig| first2 = Peter| title = Artificial Intelligence: A Modern Approach| ___location = Upper Saddle River, New Jersey| series = Prentice Hall Series in Artificial Intelligence| date = 2003}}</ref>{{rp|360–362}} A TMS explicitly tracks alternate lines of reasoning, justifications for conclusions, and lines of reasoning that lead to contradictions, allowing future reasoning to avoid these dead ends. To provide an explanation, they trace reasoning from conclusions to assumptions through rule operations or logical inferences, allowing explanations to be generated from the reasoning traces. As an example, consider a rule-based problem solver with just a few rules about Socrates that concludes he has died from poison:
{{Blockquote
|text=By just tracing through the dependency structure the problem solver can construct the following explanation: "Socrates died because he was mortal and drank poison, and all mortals die when they drink poison. Socrates was mortal because he was a man and all men are mortal. Socrates drank poison because he held dissident beliefs, the government was conservative, and those holding conservative dissident beliefs under conservative governments must drink poison."<ref name="RMS">{{Cite book| publisher = The MIT Press| isbn = 0-262-06157-0| last1 = Forbus| first1 = Kenneth| last2 = De Kleer| first2 = Johan| title = Building Problem Solvers| ___location = Cambridge, Massachusetts| date = 1993}}</ref>{{rp|164–165}}
}}
By the 1990s researchers began studying whether it is possible to meaningfully extract the non-hand-coded rules being generated by opaque trained neural networks.<ref>{{Cite journal|last1=Tickle|first1=A. B.|last2=Andrews|first2=R.|last3=Golea|first3=M.|last4=Diederich|first4=J.|date=November 1998|title=The truth will come to light: directions and challenges in extracting the knowledge embedded within trained artificial neural networks|journal=IEEE Transactions on Neural Networks|volume=9|issue=6|pages=1057–1068|doi=10.1109/72.728352|pmid=18255792|s2cid=11111578 |issn=1045-9227}}</ref> Researchers in clinical [[expert system]]s creating{{clarify|reason=who is creating? the researchers or the expert systems?|date=April 2023}} neural network-powered decision support for clinicians sought to develop dynamic explanations that allow these technologies to be more trusted and trustworthy in practice.<ref name=":0" /> In the 2010s public concerns about racial and other bias in the use of AI for criminal sentencing decisions and findings of creditworthiness may have led to increased demand for transparent artificial intelligence.<ref name="guardian" /> As a result, many academics and organizations are developing tools to help detect bias in their systems.<ref>{{cite news|url=https://www.bloomberg.com/news/articles/2018-06-13/accenture-unveils-tool-to-help-companies-insure-their-ai-is-fair|title=Accenture Unveils Tool to Help Companies Insure Their AI Is Fair|date=June 2018|work=Bloomberg.com|access-date=5 August 2018|language=en}}</ref>
[[Marvin Minsky]] et al. raised the issue that AI can function as a form of surveillance, with the biases inherent in surveillance, suggesting HI (Humanistic Intelligence) as a way to create a more fair and balanced "human-in-the-loop" AI.<ref>Minsky, et al., "The Society of Intelligent Veillance" IEEE ISTAS2013, pages 13-17.</ref>
Explainable AI has been recently a new topic researched amongst the context of modern deep learning. Modern complex AI techniques, such as [[deep learning]], are naturally opaque.<ref>{{cite magazine|last1=Mukherjee|first1=Siddhartha|title=A.I. Versus M.D.|url=https://www.newyorker.com/magazine/2017/04/03/ai-versus-md|access-date=30 January 2018|magazine=The New Yorker|date=27 March 2017}}</ref> To address this issue, methods have been developed to make new models more explainable and interpretable.<ref>{{Cite journal|date=2020-07-08|title=Interpretable neural networks based on continuous-valued logic and multicriteria decision operators|journal=Knowledge-Based Systems|language=en|volume=199|article-number=105972|doi=10.1016/j.knosys.2020.105972 |arxiv=1910.02486 |issn=0950-7051|doi-access=free|last1=Csiszár|first1=Orsolya|last2=Csiszár|first2=Gábor|last3=Dombi|first3=József}}</ref><ref name="Lipton 31–57"/><ref name="Interpretable machine learning: def"/><ref>{{cite arXiv|last1=Doshi-Velez|first1=Finale|last2=Kim|first2=Been|date=2017-02-27|title=Towards A Rigorous Science of Interpretable Machine Learning|eprint=1702.08608|class=stat.ML}}</ref><ref>{{Cite arXiv |last=Abdollahi, Behnoush, and Olfa Nasraoui.|title=Explainable Restricted Boltzmann Machines for Collaborative Filtering.|eprint=1606.07129|class=stat.ML|year=2016}}</ref><ref>{{Cite book|last1=Dombi|first1=József|last2=Csiszár|first2=Orsolya|series=Studies in Fuzziness and Soft Computing |date=2021|title=Explainable Neural Networks Based on Fuzzy Logic and Multi-criteria Decision Tools|url=https://link.springer.com/book/10.1007/978-3-030-72280-7|volume=408|language=en-gb|doi=10.1007/978-3-030-72280-7|isbn=978-3-030-72279-1|s2cid=233486978|issn=1434-9922}}</ref> This includes layerwise relevance propagation (LRP), a technique for determining which features in a particular input vector contribute most strongly to a neural network's output.<ref name="Bach Binder Montavon Klauschen p=e0130140">{{cite journal|last1=Bach|first1=Sebastian|last2=Binder|first2=Alexander|last3=Montavon|first3=Grégoire|last4=Klauschen|first4=Frederick|last5=Müller|first5=Klaus-Robert|author-link5=Klaus-Robert Müller|last6=Samek|first6=Wojciech|date=2015-07-10|editor-last=Suarez|editor-first=Oscar Deniz|title=On Pixel-Wise Explanations for Non-Linear Classifier Decisions by Layer-Wise Relevance Propagation|journal=PLOS ONE|volume=10|issue=7|page=e0130140|bibcode=2015PLoSO..1030140B|doi=10.1371/journal.pone.0130140|issn=1932-6203|pmc=4498753|pmid=26161953|doi-access=free}}</ref><ref>{{cite news|url=https://www.theguardian.com/science/2017/nov/05/computer-says-no-why-making-ais-fair-accountable-and-transparent-is-crucial|title=Computer says no: why making AIs fair, accountable and transparent is crucial|last1=Sample|first1=Ian|date=5 November 2017|work=The Guardian|access-date=5 August 2018|language=en}}</ref> Other techniques explain some particular prediction made by a (nonlinear) black-box model, a goal referred to as "local interpretability".<ref>{{Cite journal|last1=Martens|first1=David|last2=Provost|first2=Foster|title=Explaining data-driven document classifications|url=http://pages.stern.nyu.edu/~fprovost/Papers/MartensProvost_Explaining.pdf|journal=MIS Quarterly|year=2014|volume=38|pages=73–99|doi=10.25300/MISQ/2014/38.1.04|s2cid=14238842}}</ref><ref>{{Cite journal|title="Why Should I Trust You?" {{!}} Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining|language=EN|doi=10.1145/2939672.2939778|s2cid=13029170}}</ref><ref>{{Citation|last1=Lundberg|first1=Scott M|title=A Unified Approach to Interpreting Model Predictions|date=2017|url=http://papers.nips.cc/paper/7062-a-unified-approach-to-interpreting-model-predictions.pdf|work=Advances in Neural Information Processing Systems 30|pages=4765–4774|editor-last=Guyon|editor-first=I.|publisher=Curran Associates, Inc.|access-date=2020-03-13|last2=Lee|first2=Su-In|author2-link=Su-In Lee|editor2-last=Luxburg|editor2-first=U. V.|editor3-last=Bengio|editor3-first=S.|editor4-last=Wallach|editor4-first=H.|arxiv=1705.07874}}</ref><ref>{{Cite journal|last1=Carter|first1=Brandon|last2=Mueller|first2=Jonas|last3=Jain|first3=Siddhartha|last4=Gifford|first4=David|date=2019-04-11|title=What made you do this? Understanding black-box decisions with sufficient input subsets|url=http://proceedings.mlr.press/v89/carter19a.html|journal=The 22nd International Conference on Artificial Intelligence and Statistics|language=en|pages=567–576}}</ref><ref>{{Cite journal|last1=Shrikumar|first1=Avanti|last2=Greenside|first2=Peyton|last3=Kundaje|first3=Anshul|date=2017-07-17|title=Learning Important Features Through Propagating Activation Differences|url=http://proceedings.mlr.press/v70/shrikumar17a.html|journal=International Conference on Machine Learning|language=en|pages=3145–3153}}</ref><ref>{{Cite journal|url=https://dl.acm.org/doi/abs/10.5555/3305890.3306024|title=Axiomatic attribution for deep networks {{!}} Proceedings of the 34th International Conference on Machine Learning - Volume 70|website=dl.acm.org|series=Icml'17|date=6 August 2017|pages=3319–3328|language=EN|access-date=2020-03-13}}</ref> We still today cannot explain the output of today's DNNs without the new explanatory mechanisms, we also can't by the neural network, or external explanatory components <ref>{{Citation |last1=Xu |first1=Feiyu |title=Explainable AI: A Brief Survey on History, Research Areas, Approaches and Challenges |date=2019 |work=Natural Language Processing and Chinese Computing |volume=11839 |pages=563–574 |editor-last=Tang |editor-first=Jie |url=http://link.springer.com/10.1007/978-3-030-32236-6_51 |access-date=2024-12-03 |place=Cham |publisher=Springer International Publishing |language=en |doi=10.1007/978-3-030-32236-6_51 |isbn=978-3-030-32235-9 |last2=Uszkoreit |first2=Hans |last3=Du |first3=Yangzhou |last4=Fan |first4=Wei |last5=Zhao |first5=Dongyan |last6=Zhu |first6=Jun |editor2-last=Kan |editor2-first=Min-Yen |editor3-last=Zhao |editor3-first=Dongyan |editor4-last=Li |editor4-first=Sujian|url-access=subscription }}</ref> There is also research on whether the concepts of local interpretability can be applied to a remote context, where a model is operated by a third-party.<ref>{{Cite journal|last1=Le Merrer|first1=Erwan|last2=Trédan|first2=Gilles|date=September 2020|title=Remote explainability faces the bouncer problem|url=https://www.nature.com/articles/s42256-020-0216-z|journal=Nature Machine Intelligence|language=en|volume=2|issue=9|pages=529–539|doi=10.1038/s42256-020-0216-z|issn=2522-5839|arxiv=1910.01432|s2cid=225207140}}</ref><ref>{{Cite journal |last1=Aivodji |first1=Ulrich |last2=Arai |first2=Hiromi |last3=Fortineau |first3=Olivier |last4=Gambs |first4=Sébastien |last5=Hara |first5=Satoshi |last6=Tapp |first6=Alain |date=2019-05-24 |title=Fairwashing: the risk of rationalization |url=http://proceedings.mlr.press/v97/aivodji19a.html |journal=International Conference on Machine Learning |language=en |publisher=PMLR |pages=161–170 |arxiv=1901.09749}}</ref>
There has been work on making glass-box models which are more transparent to inspection.<ref name=":6"/><ref>{{cite journal |last1=Singh |first1=Chandan |last2=Nasseri |first2=Keyan |last3=Tan |first3=Yan Shuo |last4=Tang |first4=Tiffany |last5=Yu |first5=Bin |title=imodels: a python package for fitting interpretable models |journal=Journal of Open Source Software |date=4 May 2021 |volume=6 |issue=61 |pages=3192 |doi=10.21105/joss.03192 |bibcode=2021JOSS....6.3192S |s2cid=235529515 |language=en |issn=2475-9066|doi-access=free }}</ref> This includes [[decision tree]]s,<ref>{{Cite journal|last1=Vidal|first1=Thibaut|last2=Schiffer|first2=Maximilian|date=2020|title=Born-Again Tree Ensembles|url=http://proceedings.mlr.press/v119/vidal20a.html|journal=International Conference on Machine Learning|language=en|publisher=PMLR|volume=119|pages=9743–9753|arxiv=2003.11132}}</ref> [[Bayesian network]]s, sparse [[linear model]]s,<ref>{{cite journal |last1=Ustun |first1=Berk |last2=Rudin |first2=Cynthia |title=Supersparse linear integer models for optimized medical scoring systems |journal=Machine Learning |date=1 March 2016 |volume=102 |issue=3 |pages=349–391 |doi=10.1007/s10994-015-5528-6 |s2cid=207211836 |url=https://link.springer.com/article/10.1007/s10994-015-5528-6 |language=en |issn=1573-0565|arxiv=1502.04269 }}</ref> and more.<ref>Bostrom, N., & Yudkowsky, E. (2014). [https://intelligence.org/files/EthicsofAI.pdf The ethics of artificial intelligence]. ''The Cambridge Handbook of Artificial Intelligence'', 316-334.</ref> The [[ACM Conference on Fairness, Accountability, and Transparency|Association for Computing Machinery Conference on Fairness, Accountability, and Transparency (ACM FAccT)]] was established in 2018 to study transparency and explainability in the context of socio-technical systems, many of which include artificial intelligence.<ref name="FAT* conference">{{cite web | url=https://fatconference.org/ | title=FAT* Conference }}</ref><ref>{{cite news |title=Computer programs recognise white men better than black women |url=https://www.economist.com/science-and-technology/2018/02/15/computer-programs-recognise-white-men-better-than-black-women |access-date=5 August 2018 |newspaper=The Economist |date=2018 |language=en}}</ref>
Some techniques allow visualisations of the inputs to which individual [[Neuron (software)|software neurons]] respond to most strongly. Several groups found that neurons can be aggregated into circuits that perform human-comprehensible functions, some of which reliably arise across different networks trained independently.<ref name="Circuits">{{cite journal |last1=Olah |first1=Chris |last2=Cammarata |first2=Nick |last3=Schubert |first3=Ludwig |last4=Goh |first4=Gabriel |last5=Petrov |first5=Michael |last6=Carter |first6=Shan |title=Zoom In: An Introduction to Circuits |journal=Distill |date=10 March 2020 |volume=5 |issue=3 |pages=e00024.001 |doi=10.23915/distill.00024.001 |url=https://distill.pub/2020/circuits/zoom-in/ |language=en |issn=2476-0757|doi-access=free }}</ref><ref>{{cite journal |last1=Li |first1=Yixuan |last2=Yosinski |first2=Jason |last3=Clune |first3=Jeff |last4=Lipson |first4=Hod |last5=Hopcroft |first5=John |title=Convergent Learning: Do different neural networks learn the same representations? |journal=Feature Extraction: Modern Questions and Challenges |date=8 December 2015 |pages=196–212 |url=http://proceedings.mlr.press/v44/li15convergent.html |publisher=PMLR |language=en}}</ref>
There are various techniques to extract compressed representations of the features of given inputs, which can then be analysed by standard [[Cluster analysis|clustering techniques]]. Alternatively, networks can be trained to output linguistic explanations of their behaviour, which are then directly human-interpretable.<ref>{{cite book |last1=Hendricks |first1=Lisa Anne |last2=Akata |first2=Zeynep |last3=Rohrbach |first3=Marcus |last4=Donahue |first4=Jeff |last5=Schiele |first5=Bernt |last6=Darrell |first6=Trevor |title=Computer Vision – ECCV 2016 |chapter=Generating Visual Explanations |series=Lecture Notes in Computer Science |date=2016 |volume=9908 |pages=3–19 |doi=10.1007/978-3-319-46493-0_1 |chapter-url=https://link.springer.com/chapter/10.1007/978-3-319-46493-0_1 |publisher=Springer International Publishing |language=en|arxiv=1603.08507 |isbn=978-3-319-46492-3 |s2cid=12030503 }}</ref> Model behaviour can also be explained with reference to training data—for example, by evaluating which training inputs influenced a given behaviour the most,<ref>{{cite journal |last1=Koh |first1=Pang Wei |last2=Liang |first2=Percy |title=Understanding Black-box Predictions via Influence Functions |journal=International Conference on Machine Learning |date=17 July 2017 |pages=1885–1894 |url=http://proceedings.mlr.press/v70/koh17a.html |publisher=PMLR |arxiv=1703.04730 |language=en}}</ref> or by approximating its predictions using the most similar instances from the training data.<ref>{{cite journal |last1=Schmaltz |first1=Allen |title=Detecting Local Insights from Global Labels: Supervised & Zero-Shot Sequence Labeling via a Convolutional Decomposition |journal=Computational Linguistics |date=23 December 2021 |volume=47 |issue=4 |pages=729–773 |doi=10.1162/coli_a_00416 |doi-access=free }}</ref>
The use of explainable artificial intelligence (XAI) in pain research, specifically in understanding the role of electrodermal activity for [[Automated Pain Recognition|automated pain recognition]]: hand-crafted features and deep learning models in pain recognition, highlighting the insights that simple hand-crafted features can yield comparative performances to deep learning models and that both traditional feature engineering and deep feature learning approaches rely on simple characteristics of the input time-series data.<ref>{{Cite journal |last1=Gouverneur |first1=Philip |last2=Li |first2=Frédéric |last3=Shirahama |first3=Kimiaki |last4=Luebke |first4=Luisa |last5=Adamczyk |first5=Wacław M. |last6=Szikszay |first6=Tibor M. |last7=Luedtke |first7=Kerstin |last8=Grzegorzek |first8=Marcin |date=2023-02-09 |title=Explainable Artificial Intelligence (XAI) in Pain Research: Understanding the Role of Electrodermal Activity for Automated Pain Recognition |journal=Sensors |language=en |volume=23 |issue=4 |pages=1959 |doi=10.3390/s23041959 |issn=1424-8220 |pmc=9960387 |pmid=36850556 |bibcode=2023Senso..23.1959G |doi-access=free }}</ref>
== Regulation ==
As regulators, official bodies, and general users come to depend on AI-based dynamic systems, clearer accountability will be required for [[automated decision-making]] processes to ensure trust and transparency. The first global conference exclusively dedicated to this emerging discipline was the 2017 [[International Joint Conference on Artificial Intelligence]]: Workshop on Explainable Artificial Intelligence (XAI).<ref>{{cite web|title=IJCAI 2017 Workshop on Explainable Artificial Intelligence (XAI)|url=http://www.intelligentrobots.org/files/IJCAI2017/IJCAI-17_XAI_WS_Proceedings.pdf|website=Earthlink|publisher=IJCAI|access-date=17 July 2017|archive-date=4 April 2019|archive-url=https://web.archive.org/web/20190404131609/http://www.intelligentrobots.org/files/IJCAI2017/IJCAI-17_XAI_WS_Proceedings.pdf|url-status=dead}}</ref> It has evolved over the years, with various workshops organised and co-located to many other international conferences, and it has now a dedicated global event, "The world conference on eXplainable Artificial Intelligence", with its own proceedings.<ref name="XAI-2023">{{cite book |author=<!--Not stated--> |date= 2023| title= Explainable Artificial Intelligence, First World Conference, xAI 2023, Lisbon, Portugal, July 26–28, 2023, Proceedings, Parts I/II/III |series= Communications in Computer and Information Science|volume= 1903|url= https://link.springer.com/book/10.1007/978-3-031-44070-0 |publisher=springer |doi= 10.1007/978-3-031-44070-0|isbn=978-3-031-44070-0}}</ref><ref name="XAI-2024">{{cite book |author=<!--Not stated-->|date= 2024| title= Explainable Artificial Intelligence, Second World Conference, xAI 2024, Valletta, Malta, July 17–19, 2024, Proceedings, Part I/II/III/IV |series= Communications in Computer and Information Science|volume= 2153|url=https://link.springer.com/book/10.1007/978-3-031-63787-2 |publisher=springer |doi= 10.1007/978-3-031-63787-2|isbn=978-3-031-63787-2}}</ref>
The European Union introduced a [[right to explanation]] in the [[General Data Protection Regulation]] (GDPR) to address potential problems stemming from the rising importance of algorithms. The implementation of the regulation began in 2018. However, the right to explanation in GDPR covers only the local aspect of interpretability. In the United States, insurance companies are required to be able to explain their rate and coverage decisions.<ref>{{cite news |last1=Kahn |first1=Jeremy |title=Artificial Intelligence Has Some Explaining to Do |url=https://www.bloomberg.com/news/articles/2018-12-12/artificial-intelligence-has-some-explaining-to-do |access-date=17 December 2018 |work=[[Bloomberg Businessweek]] |date=12 December 2018}}</ref> In France the [[Loi pour une République numérique]] (Digital Republic Act) grants subjects the right to request and receive information pertaining to the implementation of algorithms that process data about them.
== Limitations ==
Despite ongoing endeavors to enhance the explainability of AI models, they persist with several inherent limitations.
=== Adversarial parties ===
By making an AI system more explainable, we also reveal more of its inner workings. For example, the explainability method of feature importance identifies features or variables that are most important in determining the model's output, while the influential samples method identifies the training samples that are most influential in determining the output, given a particular input.<ref name="Explainable Machine Learning in Deployment">{{cite book | last1=Bhatt | first1=Umang | last2=Xiang | first2=Alice | last3=Sharma | first3=Shubham | last4=Weller | first4=Adrian | last5=Taly | first5=Ankur | last6=Jia | first6=Yunhan | last7=Ghosh | first7=Joydeep | last8=Puri | first8=Richir | last9=M.F. Moura | first9=José | last10=Eckersley | first10=Peter | title=Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency | chapter=Explainable Machine Learning in Deployment | date=2022 | pages=648–657 | doi=10.1145/3351095.3375624 | isbn=9781450369367 | s2cid=202572724 | chapter-url=https://dl.acm.org/doi/pdf/10.1145/3351095.3375624 }}</ref> Adversarial parties could take advantage of this knowledge.
For example, competitor firms could replicate aspects of the original AI system in their own product, thus reducing competitive advantage.<ref name="How the machine 'thinks'">{{cite journal |last1=Burrel |first1=Jenna |date=2016 |title=How the machine 'thinks': Understanding opacity in machine learning algorithms |journal=Big Data & Society |volume=3 |issue=1 |article-number=2053951715622512 |doi=10.1177/2053951715622512 |s2cid=61330970|doi-access=free }}</ref> An explainable AI system is also susceptible to being “gamed”—influenced in a way that undermines its intended purpose. One study gives the example of a predictive policing system; in this case, those who could potentially “game” the system are the criminals subject to the system's decisions. In this study, developers of the system discussed the issue of criminal gangs looking to illegally obtain passports, and they expressed concerns that, if given an idea of what factors might trigger an alert in the passport application process, those gangs would be able to “send guinea pigs” to test those triggers, eventually finding a loophole that would allow them to “reliably get passports from under the noses of the authorities”.<ref>{{cite book | last1=Veale | first1=Michael | last2=Van Kleek | first2=Max | last3=Binns | first3=Reuben | title=Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems | chapter=Fairness and Accountability Design Needs for Algorithmic Support in High-Stakes Public Sector Decision-Making | date=2018 | volume=40 | pages=1–14 | doi=10.1145/3173574.3174014 | isbn=9781450356206 | s2cid=3639135 | chapter-url=https://dl.acm.org/doi/pdf/10.1145/3173574.3174014 }}</ref>
=== Adaptive integration and explanation ===
Many approaches that it uses provides explanation in general, it doesn't take account for the diverse backgrounds and knowledge level of the users. This leads to challenges with accurate comprehension for all users. Expert users can find the explanations lacking in depth, and are oversimplified, while a beginner user may struggle understanding the explanations as they are complex. This limitation downplays the ability of the XAI techniques to appeal to their users with different levels of knowledge, which can impact the trust from users and who uses it. The quality of explanations can be different amongst their users as they all have different expertise levels, including different situation and conditions.<ref>{{Cite journal |last1=Yang |first1=Wenli |last2=Wei |first2=Yuchen |last3=Wei |first3=Hanyu |last4=Chen |first4=Yanyu |last5=Huang |first5=Guan |last6=Li |first6=Xiang |last7=Li |first7=Renjie |last8=Yao |first8=Naimeng |last9=Wang |first9=Xinyi |last10=Gu |first10=Xiaotong |last11=Amin |first11=Muhammad Bilal |last12=Kang |first12=Byeong |date=2023-08-10 |title=Survey on Explainable AI: From Approaches, Limitations and Applications Aspects |journal=Human-Centric Intelligent Systems |language=en |volume=3 |issue=3 |pages=161–188 |doi=10.1007/s44230-023-00038-y |issn=2667-1336|doi-access=free }}</ref>
=== Technical complexity ===
A fundamental barrier to making AI systems explainable is the technical complexity of such systems. End users often lack the coding knowledge required to understand software of any kind. Current methods used to explain AI are mainly technical ones, geared toward machine learning engineers for debugging purposes, rather than toward the end users who are ultimately affected by the system, causing “a gap between explainability in practice and the goal of transparency”.<ref name="Explainable Machine Learning in Deployment" /> Proposed solutions to address the issue of technical complexity include either promoting the coding education of the general public so technical explanations would be more accessible to end users, or providing explanations in layperson terms.<ref name="How the machine 'thinks'" />
The solution must avoid oversimplification. It is important to strike a balance between accuracy – how faithfully the explanation reflects the process of the AI system – and explainability – how well end users understand the process. This is a difficult balance to strike, since the complexity of machine learning makes it difficult for even ML engineers to fully understand, let alone non-experts.<ref name="Explainable Machine Learning in Deployment" />
=== Understanding versus trust ===
The goal of explainability to end users of AI systems is to increase trust in the systems, even “address concerns about lack of ‘fairness’ and discriminatory effects”.<ref name="How the machine 'thinks'" /> However, even with a good understanding of an AI system, end users may not necessarily trust the system.<ref>{{cite journal |last1=Hu |first1=Tongxi |last2=Zhang |first2=Xuesong |last3=Bohrer |first3=Gil |last4=Liu |first4=Yanlan |last5=Zhou |first5=Yuyu |last6=Martin |first6=Jay |last7=LI |first7=Yang |last8=Zhao |first8=Kaiguang |title=Crop yield prediction via explainable AI and interpretable machine learning: Dangers of black box models for evaluating climate change impacts on crop yield|journal=Agricultural and Forest Meteorology |date=2023 |volume=336 |article-number=109458 |doi=10.1016/j.agrformet.2023.109458 |bibcode=2023AgFM..33609458H |s2cid=258552400 |doi-access=free }}</ref> In one study, participants were presented with combinations of white-box and black-box explanations, and static and interactive explanations of AI systems. While these explanations served to increase both their self-reported and objective understanding, it had no impact on their level of trust, which remained skeptical.<ref name="Explaining Decision-Making Algorithms through UI">{{cite book | last1=Cheng | first1=Hao-Fei | last2=Wang | first2=Ruotang | last3=Zhang | first3=Zheng | last4=O’Connell | first4=Fiona | last5=Gray | first5=Terrance | last6=Harper | first6= F. Maxwell | last7=Zhu | first7=Haiyi | title=Explaining Decision-Making Algorithms through UI: Strategies to Help Non-Expert Stakeholders | series=Proceedings of the 2019 CHI Conference on Human Factors in Computing Systems | date=2019 | volume=559 | pages=1–12 | doi=10.1145/3290605.3300789 | isbn=9781450359702 | s2cid=140281803 | url=https://dl.acm.org/doi/pdf/10.1145/3290605.3300789 }}</ref>
This outcome was especially true for decisions that impacted the end user in a significant way, such as graduate school admissions. Participants judged algorithms to be too inflexible and unforgiving in comparison to human decision-makers; instead of rigidly adhering to a set of rules, humans are able to consider exceptional cases as well as appeals to their initial decision.<ref name="Explaining Decision-Making Algorithms through UI" /> For such decisions, explainability will not necessarily cause end users to accept the use of decision-making algorithms. We will need to either turn to another method to increase trust and acceptance of decision-making algorithms, or question the need to rely solely on AI for such impactful decisions in the first place.
However, some emphasize that the purpose of explainability of artificial intelligence is not to merely increase users' trust in the system's decisions, but to calibrate the users' level of trust to the correct level.<ref>{{Cite journal |last1=Liel |first1=Yotam |last2=Zalmanson |first2=Lior |date=August 2023 |title=Turning Off Your Better Judgment – Conformity to Algorithmic Recommendations |url=http://journals.aom.org/doi/full/10.5465/AMPROC.2023.277bp |journal=Academy of Management Proceedings |language=en |volume=2023 |issue=1 |article-number=16319 |doi=10.5465/AMPROC.2023.277bp |s2cid=260212999 |issn=0065-0668|url-access=subscription }}</ref> According to this principle, too much or too little user trust in the AI system will harm the overall performance of the human-system unit. When the trust is excessive, the users are not critical of possible mistakes of the system and when the users do not have enough trust in the system, they will not exhaust the benefits inherent in it.
== Criticism ==
Some scholars have suggested that explainability in AI should be considered a goal secondary to AI effectiveness, and that encouraging the exclusive development of XAI may limit the functionality of AI more broadly.<ref name=":5">{{Cite journal|last1=McCoy|first1=Liam G.|last2=Brenna|first2=Connor T. A.|last3=Chen|first3=Stacy S.|last4=Vold|first4=Karina|last5=Das|first5=Sunit|date=2021-11-05|title=Believing in black boxes: machine learning for healthcare does not need explainability to be evidence-based|url=https://www.jclinepi.com/article/S0895-4356(21)00354-1/abstract|journal=Journal of Clinical Epidemiology|volume=142|language=English|issue=Online ahead of print|pages=252–257|doi=10.1016/j.jclinepi.2021.11.001|issn=0895-4356|pmid=34748907|s2cid=243810442}}</ref><ref>{{Cite journal|last1=Ghassemi|first1=Marzyeh|last2=Oakden-Rayner|first2=Luke|last3=Beam|first3=Andrew L.|date=2021-11-01|title=The false hope of current approaches to explainable artificial intelligence in health care|journal=The Lancet Digital Health|language=English|volume=3|issue=11|pages=e745–e750|doi=10.1016/S2589-7500(21)00208-9|issn=2589-7500|pmid=34711379|s2cid=239963176|doi-access=free}}</ref> Critiques of XAI rely on developed concepts of mechanistic and empiric reasoning from [[evidence-based medicine]] to suggest that AI technologies can be clinically validated even when their function cannot be understood by their operators.<ref name=":5"/>
Some researchers advocate the use of inherently interpretable machine learning models, rather than using post-hoc explanations in which a second model is created to explain the first. This is partly because post-hoc models increase the complexity in a decision pathway and partly because it is often unclear how faithfully a post-hoc explanation can mimic the computations of an entirely separate model.<ref name=":6" /> However, another view is that what is important is that the explanation accomplishes the given task at hand, and whether it is pre or post-hoc doesn't matter. If a post-hoc explanation method helps a doctor diagnose cancer better, it is of secondary importance whether it is a correct/incorrect explanation.
The goals of XAI amount to a form of [[Lossy compression artefact|lossy compression]] that will become less effective as AI models grow in their number of parameters. Along with other factors this leads to a theoretical limit for explainability.<ref>{{Cite journal |last=Sarkar |first=Advait |date=2022 |title=Is explainable AI a race against model complexity? |url=http://ceur-ws.org/Vol-3124/paper22.pdf |journal=Workshop on Transparency and Explanations in Smart Systems (TeXSS), in Conjunction with ACM Intelligent User Interfaces (IUI 2022) |pages=192–199 |arxiv=2205.10119 |via=CEUR Workshop Proceedings}}</ref>
== Explainability in social choice ==
Explainability was studied also in [[social choice theory]]. Social choice theory aims at finding solutions to social decision problems, that are based on well-established axioms. [[Ariel D. Procaccia]]<ref>{{Citation |last=Procaccia |first=Ariel D. |title=Axioms Should Explain Solutions |date=2019 |work=The Future of Economic Design: The Continuing Development of a Field as Envisioned by Its Researchers |series=Studies in Economic Design |pages=195–199 |editor-last=Laslier |editor-first=Jean-François |url=https://doi.org/10.1007/978-3-030-18050-8_27 |access-date=2024-03-21 |place=Cham |publisher=Springer International Publishing |language=en |doi=10.1007/978-3-030-18050-8_27 |isbn=978-3-030-18050-8 |editor2-last=Moulin |editor2-first=Hervé |editor3-last=Sanver |editor3-first=M. Remzi |editor4-last=Zwicker |editor4-first=William S.|url-access=subscription }}</ref> explains that these axioms can be used to construct convincing explanations to the solutions. This principle has been used to construct explanations in various subfields of social choice.
=== Voting ===
Cailloux and Endriss<ref>{{Cite journal |last1=Cailloux |first1=Olivier |last2=Endriss |first2=Ulle |date=2016-05-09 |title=Arguing about Voting Rules |url=https://dl.acm.org/doi/abs/10.5555/2936924.2936968 |journal=Proceedings of the 2016 International Conference on Autonomous Agents & Multiagent Systems |series=AAMAS '16 |___location=Richland, SC |publisher=International Foundation for Autonomous Agents and Multiagent Systems |pages=287–295 |isbn=978-1-4503-4239-1}}</ref> present a method for explaining voting rules using the [[axiom]]s that characterize them. They exemplify their method on the [[Borda voting|Borda voting rule]] .
Peters, Procaccia, Psomas and Zhou<ref>{{Cite journal |last1=Peters |first1=Dominik |last2=Procaccia |first2=Ariel D |last3=Psomas |first3=Alexandros |last4=Zhou |first4=Zixin |date=2020 |title=Explainable Voting |url=https://proceedings.neurips.cc/paper/2020/hash/10c72a9d42dd07a028ee910f7854da5d-Abstract.html |journal=Advances in Neural Information Processing Systems |publisher=Curran Associates, Inc. |volume=33 |pages=1525–1534}}</ref> present an algorithm for explaining the outcomes of the Borda rule using O(''m''<sup>2</sup>) explanations, and prove that this is tight in the worst case.
=== Participatory budgeting ===
Yang, Hausladen, Peters, Pournaras, Fricker and Helbing<ref name=":12">{{Cite journal |arxiv=2310.03501 |first1=Joshua C. |last1=Yang |first2=Carina I. |last2=Hausladen |title=Designing Digital Voting Systems for Citizens: Achieving Fairness and Legitimacy in Participatory Budgeting |date=2024 |last3=Peters |first3=Dominik |last4=Pournaras |first4=Evangelos |author5=Regula Häenggli Fricker |last6=Helbing |first6=Dirk|journal=Digital Government: Research and Practice |volume=5 |issue=3 |pages=1–30 |doi=10.1145/3665332 }}</ref> present an empirical study of explainability in [[participatory budgeting]]. They compared the greedy and the [[Method of equal shares|equal shares]] rules, and three types of explanations: ''mechanism explanation'' (a general explanation of how the aggregation rule works given the voting input), ''individual explanation'' (explaining how many voters had at least one approved project, at least 10000 CHF in approved projects), and ''group explanation'' (explaining how the budget is distributed among the districts and topics). They compared the perceived ''trustworthiness'' and ''fairness'' of greedy and equal shares, before and after the explanations. They found out that, for MES, mechanism explanation yields the highest increase in perceived fairness and trustworthiness; the second-highest was Group explanation. For Greedy, Mechanism explanation increases perceived trustworthiness but not fairness, whereas Individual explanation increases both perceived fairness and trustworthiness. Group explanation ''decreases'' the perceived fairness and trustworthiness.
=== Payoff allocation ===
Nizri, Azaria and Hazon<ref>{{Cite journal |last1=Nizri |first1=Meir |last2=Hazon |first2=Noam |last3=Azaria |first3=Amos |date=2022-06-28 |title=Explainable Shapley-Based Allocation (Student Abstract) |url=https://ojs.aaai.org/index.php/AAAI/article/view/21648 |journal=Proceedings of the AAAI Conference on Artificial Intelligence |language=en |volume=36 |issue=11 |pages=13023–13024 |doi=10.1609/aaai.v36i11.21648 |s2cid=250296641 |issn=2374-3468|doi-access=free }}</ref> present an algorithm for computing explanations for the [[Shapley value]]. Given a coalitional game, their algorithm decomposes it to sub-games, for which it is easy to generate verbal explanations based on the axioms characterizing the Shapley value. The payoff allocation for each sub-game is perceived as fair, so the Shapley-based payoff allocation for the given game should seem fair as well. An experiment with 210 human subjects shows that, with their automatically generated explanations, subjects perceive Shapley-based payoff allocation as significantly fairer than with a general standard explanation.
== See also ==
* {{annotated link|Algorithmic transparency}}
* {{annotated link|Right to explanation}}
* {{annotated link|Accumulated local effects}}
== References ==
{{
== External links ==
* {{ cite web | url=http://xaiworldconference.com/ | title=the World Conference on eXplainable Artificial Intelligence}}
* {{cite web | url=https://fatconference.org/ | title=ACM Conference on Fairness, Accountability, and Transparency (FAccT) }}
* {{ Cite journal | title= Random Forest similarity maps: A Scalable Visual Representation for Global and Local Interpretation| year= 2021| doi= 10.3390/electronics10222862| doi-access= free| last1= Mazumdar| first1= Dipankar| last2= Neto| first2= Mário Popolin| last3= Paulovich| first3= Fernando V.| journal= Electronics| volume= 10| issue= 22| page= 2862}}
* {{cite web | title=Explainable AI: Making machines understandable for humans | website=Explainable AI: Making machines understandable for humans | url=https://explainableai.com/ | ref={{sfnref | Explainable AI: Making machines understandable for humans}} | access-date=2017-11-02}}
* {{cite web |title=Explaining How End-to-End Deep Learning Steers a Self-Driving Car |website=Parallel Forall |date=2017-05-23 |url=https://devblogs.nvidia.com/parallelforall/explaining-deep-learning-self-driving-car/ |ref={{sfnref | Parallel Forall | 2017}} |access-date=2017-11-02}}
* {{cite web | last=Knight | first=Will | title=DARPA is funding projects that will try to open up AI's black boxes | website=MIT Technology Review | date=2017-03-14 | url=https://www.technologyreview.com/s/603795/the-us-military-wants-its-autonomous-machines-to-explain-themselves/ | access-date=2017-11-02}}
* {{cite arXiv | last1=Alvarez-Melis | first1=David | last2=Jaakkola | first2=Tommi S. |title=A causal framework for explaining the predictions of black-box sequence-to-sequence models | date=2017-07-06 | eprint=1707.01943 | class=cs.LG }}
* {{cite web | title=Similarity Cracks the Code Of Explainable AI | website=simMachines | date=2017-10-12 | url=http://simmachines.com/similarity-cracks-code-explainable-ai/ | ref={{sfnref | simMachines | 2017}} | access-date=2018-02-02}}
{{Differentiable computing}}
{{DEFAULTSORT:Explainable AI}}
[[Category:Quality control tools]]
[[Category:Artificial intelligence]]
[[Category:
| |||