(WIP) Reliable IA

Table of Contents

• (WIP) Reliable IA
  • Table of Contents
  • Contributing
• Introduction : why
• Scientific Themes
  • Transparency
  • Ethics
  • Robutness
  • Environment
  • Reproducibility
• Platform :
• Stategic Themes
  • Human-Centered: "Contrôle Humain"
  • Responsability
  • Privacy

Contributing

Please feel free to send me pull requests or email ([email protected]) to add links.

Introduction : why

Notes :

• Responsable vs Irresponsable (= ne pas avoir conséquence de ses actes)

• Responsible AI : 6 keys (Human-Centered ML, Secure, Interpretable IA, Explainable, Ethics, Compliance)

• A LIRE:

  • https://betaandbit.github.io/RML/#p=1
  • []: https://www.datanami.com/2020/04/06/brief-perspective-on-key-terms-and-ideas-in-responsible-ai/ Thrustworthy/Reliable AI : 7 keys Scientific (Human-Centered ML, Robustness, Ethics, Environmental, Transparency,) & Strategic (Privacy & Data Gouv, Tracking & Reproducible operations)
  • []: https://www.emedgene.com/7-keys-to-a-trustworthy-ai-according-to-the-eu-guidelines/
  • https://docs.google.com/presentation/d/1Md24K25opDU9lb5llop8i_vYs1aLvryW9iemF1y6gAU/edit#slide=id.p33
  • CausualML Challenge: https://neurips.cc/Conferences/2022/CompetitionTrack

• Ex. AI fails :

  • Understand how it works ? « Ballon foot », « Barbe masque détection »
  • Impact société dû biais : Recrutement, COMPAS, etc.
  • AutoML - performance basée sur des métriques techniques (eg. accuracy) et non business (e.g. recrutement)

• Législation :

  • Avant sur la Data : RGPD (2016)
  • Maintient sur l’IA : Thrustworthy IA, DI US
  • Guide Pratique : https://www.afjv.com/news/10981_guide-pratique-nouveau-reglement-ia.htm
  • OECD: https://oecd.ai/en/accountability

• Responsible ML :

  • https://info.h2o.ai/rs/644-PKX-778/images/OReilly_Responsible_ML_eBook.pdf

• CNIL

  • https://www.cnil.fr/fr/intelligence-artificielle/guide

Scientific Themes

• Monitoring

Model observability

• https://arize.com/ml-observability/

• https://neptune.ai/blog/ml-model-monitoring-best-tools

• https://towardsdatascience.com/ml-infrastructure-tools-ml-observability-8e4d7df6db43

• https://towardsdatascience.com/what-is-ml-observability-29e85e701688

• https://medium.com/arize-ai/ml-infrastructure-tools-ml-observability-4b74d05a5fd6

• https://www.montecarlodata.com/blog-beyond-monitoring-the-rise-of-observability/

• Alerting

Transparency

• Methodology
• Explainability
• Interpretable IA

Questions:

• How does each features contribute to a model's prediction?
• How does a prediction change dependent on feature inputs?
• What features are or not are significant for a given outcome?
• What features would you change to obtain a different prediction?
• How robust is the model?

Tools:

• https://github.com/SelfExplainML/PiML-Toolbox
• https://github.com/g8a9/ferret
• explainerdashboard

Shap: Patrick Hall

• https://maven.com/p/d3409b/understand-shap-s-hapley-additive-ex-planations

A Selection of Medium articles :

• https://pub.towardsai.net/shapash-making-ml-models-understandable-by-everyone-8f96ad469eb3
• https://www.marktechpost.com/2022/02/10/uc-berkeley-researchers-introduce-imodels-a-python-package-for-fitting-interpretable-machine-learning-models/
• LimeCraft: https://arxiv.org/pdf/2111.08094.pdf
• Xplique: https://github.com/deel-ai/xplique

• Shap (Shapley Additive exPlanations) : Shap is a model agnostic and works by breaking down the contribution of each feature and attributing a score to each feature.

• LIME (local Interpretable Model-agnostic Explanations) : LIME is another model agnostic method that works by approximinating the behavior of the model locally around a specific prediction.

• Eli5 : library for debugging and explaining classifiers. It provides feature importance scores, as well as "reason codes" for scikit-learn, Keras, Xgboost, LightGBM, CatBoost

• Shapash : python library which aims to make machine learning interpretable and understandable to everyone. Shapash provides several types of visualization with explicit labels.

• Anchors : method for generating humain-interpretable rules that can be used to explain the predictions of a machine learning model.

• XAI (eXplainable AI): XAI is a library for explaining and visualizeing the predictions of machine learning models including feature importance scores, decision trees, and rule-based explanations.

• BreakDown : tool that can be used to explain the predictions of linear models. It works by decomposing the model's output into the contribution of each input feature.

• Interpret-text: interpret-text is a library for explaining the predictions of natural language processing models.

• iml (Interpretable MAchine Learning): iml currently contains the interface and IO code from the Shap project and it will potentially also do the same for the Lime project.

• aix360 (AI Explainibility 360): aix360 includes a comprehensive set of algorithms that cover different dimensions.

• OmnniXAI (short for Omni eXplainable AI), adresses several problems with interpreting judgements produced by machine learning models in practice.

• Seldon (alibi explain / detect):

  • https://www.seldon.io/solutions/open-source-projects/alibi-explain
  • https://www.seldon.io/solutions/open-source-projects/alibi-detect

• Source: https://www.linkedin.com/posts/maryammiradi_ai-explainability-beyond-the-surface-uncommon-activity-7212004104263380993-qmcn/?utm_source=share&utm_medium=member_ios

Ethics

• https://gendered-news.imag.fr/genderednews/

• Methodology

• Lesson :

  • https://ethics.fast.ai/
  • https://ethical.institute/index.html

• https://www.youtube.com/watch?v=0Q4wU2dyMbI&t=161s

• https://github.com/dssg/aequitas

• Bais:

  • https://causalnex.readthedocs.io/en/latest/03_tutorial/04_sklearn_tutorial.html#Dataset-bias-evaluation
  • Comment mesurer les biais dans les données ? https://www.youtube.com/watch?v=2df7doSlUwA

• Identifying and managing bias in AI: https://doi.org/10.6028/NIST.SP.1270 / https://nvlpubs.nist.gov/nistpubs/SpecialPublications/NIST.SP.1270.pdf

• DALLE: https://www.vox.com/future-perfect/23023538/ai-dalle-2-openai-bias-gpt-3-incentives

• Google: https://ai.googleblog.com/2018/09/introducing-inclusive-images-competition.html

• https://github.com/Trusted-AI/AIF360/tree/master/examples

• Facebook, Balance : https://github.com/facebookresearch/balance

Fairness

• https://arxiv.org/pdf/2102.08453.pdf

Robutness

• Notes

• http://www.trustworthymachinelearning.com/

• Ex.
  • Confiance interval
  • Perturbations : graines aléatoire
  • MAPIE

Environment

• Impact environment : carbonml
• Solutions : small data ?

Reproducibility

Notes:

• Tracking
• Model registry
• Train Dataset

Tools:

• mlflow

Platform :

• https://github.com/Giskard-AI/giskard

Stategic Themes

Human-Centered: "Contrôle Humain"

Responsability

Privacy

Quelques sources:

• Algo Audit: http://algaudit.inrialpes.fr/
• Projet INRIA régulation numérique: https://www.inria.fr/en/regalia-pilot-project-regulation-algorithms
• Interview Clément Henin et Daniel Le Métayer: https://linc.cnil.fr/fr/clement-henin-et-daniel-le-metayer-fournir-des-explications-du-fonctionnement-des-algorithmes
• Vidéo BigData Paris: https://www.alain-bensoussan.com/avocats/nouveau-reglement-sur-lia-pour-une-ia-digne-de-confiance/2021/10/25/
• Réglement: https://www.senat.fr/europe/textes_europeens/COM_2021_206.pdf
• Human-Learn: https://github.com/koaning/human-learn
• Labelia Labs: https://github.com/LabeliaLabs/referentiel-evaluation-dsrc
• Méthodo : https://dataanalyticspost.com/grille-evaluation-dispositifs-medicaux/amp/
• Fiancial Risk Management and Explainable Trusworthy, Responsible AI: https://www.frontiersin.org/articles/10.3389/frai.2022.779799/full
• "Domaine de validité": https://www.quantmetry.com/blog/domaine-de-validite-ia-confiance/
• Faire émerger un cadre sur IA Confiance (construction outil IA de Confiance - Responsable (label IA resp en FR et Europe de l'Ouest): https://www.youtube.com/watch?v=Ip4dCZ8xhEo
• Implicity: autorisation FDA, algo ECG: https://www.prnewswire.com/news-releases/implicity-receives-fda-clearance-for-ai-powered-ecg-analyzer-for-implantable-loop-recorders-301446711.html?tc=eml_cleartime
• https://fortune-com.cdn.ampproject.org/c/s/fortune.com/2022/03/22/ai-explainable-radiology-medicine-crisis-eye-on-ai/amp/
• ML in High-Risk Applications: https://learning.oreilly.com/library/view/machine-learning-for/9781098102425/
• https://www.seldon.io/using-explainable-ai-xai-for-compliance-and-trust-in-the-healthcare-industry
• Trusworthy AI: https://csdl-downloads.ieeecomputer.org/mags/co/2023/02/10042078.pdf?Expires=1677230283&Policy=eyJTdGF0ZW1lbnQiOlt7IlJlc291cmNlIjoiaHR0cHM6Ly9jc2RsLWRvd25sb2Fkcy5pZWVlY29tcHV0ZXIub3JnL21hZ3MvY28vMjAyMy8wMi8xMDA0MjA3OC5wZGYiLCJDb25kaXRpb24iOnsiRGF0ZUxlc3NUaGFuIjp7IkFXUzpFcG9jaFRpbWUiOjE2NzcyMzAyODN9fX1dfQ__&Signature=bTlcKKXtID1zywcPUxJSfte2GKSLWwKYxaUZb53hMCrbcoholFbfKys5nAv-qDwJXTpFFd0JXj~s0FH0sx9IDfXNcEocUFVBmJcaoy17YqWlPtjuG9QihbwZsl0qkcRdnMHrbLB7n5fl1yDO17aAl0d2qzCkpmH8XYQnytgvuCMka2jGqdUEnAvl8EgW3hQMB6oyvOc2dw-ndBoaVJJssvqqt7Dw~qmKlyVuCOX48VKmM5LP8ear1ZCtbn1fgU87ZaDIRj3XuiOsqZUYCRcpaPABFOr3oK~z3Y4~GdbFntjhf7J8JB80elaO15RaE487SMkeGaYq6vKJVlGLJTn-SA__&Key-Pair-Id=K12PMWTCQBDMDT
• AAAI Spring Symposium 2023: https://aita.sciencesconf.org/
• Example of ethical charter (pole emploi): https://www.pole-emploi.org/files/live/sites/peorg/files/images/Communiqu%c3%a9%20de%20presse/Charte%20de%20p%c3%b4le%20emploi%20pour%20une%20Intelligence%20Artificielle%20%c3%a9....pdf

Books:

• ML in High-Risk Applications: https://learning.oreilly.com/library/view/machine-learning-for/9781098102425/
  • Chp1: Contemporary Model Governance: "Going fast and breaking thinkgs. It can mean that a small group of data scientists and engineers causes real harm at scale to many people." -> Cas d'application sur la voiture autonome chez Uber (gestion incidents, risk management, documentation).
  • Chap2: Debugging ML Systems: "Tests data area under the curve (AUC) tells us almost nothing aboout harms or security vulnerabilities. Yet these problems are often whu AI systems fail once deployed." -> Cas d'application octroi de crédit (détection de dérives, stress-tests).
  • Chap3: Security for ML: "The worst ennemy of security is complexity. Unduly complex AI systems are innately insecure." -> Censure anti-terroriste de FB (Attaques, vol de données / modèles, sécurité IT)

• https://csdl-downloads.ieeecomputer.org/mags/co/2023/02/10042078.pdf?Expires=16775762de%20l%27orga%20hi%C3%A9rarchique39&Policy=eyJTdGF0ZW1lbnQiOlt7IlJlc291cmNlIjoiaHR0cHM6Ly9jc2RsLWRvd25sb2Fkcy5pZWVlY29tcHV0ZXIub3JnL21hZ3MvY28vMjAyMy8wMi8xMDA0MjA3OC5wZGYiLCJDb25kaXRpb24iOnsiRGF0ZUxlc3NUaGFuIjp7IkFXUzpFcG9jaFRpbWUiOjE2Nzc1NzYyMzl9fX1dfQ__&Signature=q4XZkM-AYoMgGHBnRE9iIJlIWMA5BCQBTTqtdftLtIaXuYRy1-eLUdzP4Vz1yPgu2hm8FBcKS3-Fx-fvR1p61ur7ehV2-I5jLjPERv5kZuLDfXvX91G566akf8kbaudTzUY37cNhC5-Dvl8kxiN274fh1PwrindCf2a-SJWkNwH8BMTmnyNEUjsGz2PMIIBsAyROSg3wTHL2pS6l0Rz6cMC65BFRwTbw0Dp3uj3Q2R1Vz~5k-P7Lcj03ETpyDZ-DoIA28cVfyCRIUXcd0HRo~Qx24iZtnhuo8QDzXnrMTJRm6X4G7qiZB9I4-OMz43nmsWpTm0jGK1xq1bnNIPqtYA__&Key-Pair-Id=K12PMWTCQBDMDT