The common good at the risk of algorithms What to do with AI? April 9th2021 E. de Rocquigny - ERIAC
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
E. de Rocquigny MINI-CV President of OperationData Inc. and advisor of half a dozen algorithmic companies, founder think-tank Espérance & Algorithmes, Form er Full Profes s or in Applied Mathem atics & Deputy Dean of Res earch at the Ecole Centrale Paris -Univers ité Paris -Saclay, Form er Senior Scientis t EDF R&D, Expert at the National Res earch Agency, European Com m is s ion, Uni Bocconi. Links ● How to reinvent your business with data andAI : MOOC avec Bpifrance (in French) ● Managing uncertainty & creating value with robust algorithms : Reference Book ● Manifes to www.es perance-algorithmes .org edr@operationdata.fr
Plan Algorithms - what are we talking about? The need for algorithmic betting Discernment: conditions for well-balanced algorithmic services Real-world tensions, levels of stakes - creative compromise Employment and AI
Universal digital machines ◆Sales/customer experience Wine, parking ◆Innovation/product design manufacturing ◆production Co-operatives, human services, recruitment ◆Supply Chain Construction, car fleet ◆Control/ back office Accounting, law, business services ◆...
Scalability & monopolies Automation - optimization (radical) Turnover customization (massive) Operating expenses Servicification (dominant)
To the common good - with all our intelligence Toda y I ta ke hea ven a nd ea rth a s a witnes s a ga ins t you: I put before you life or dea th, If nothing should be done for the certain, (...) bles s ing or curs e. So choos e life, s o tha t nothing should be done at all because nothing is you a nd your des cenda nts m a y live,... certain (...) But when we work for tomorrow and By loving the Lord your God, lis tening to his for the uncertain, we act with reason: because we voice, a tta ching yours elf to Him ; this is must work for the uncertain by the rule of parties where your life is . - Deuteronomy 30: 19-20 that is demonstrated. - Pascal, Pensées Indeed, creation eagerly awaits the revelation of the sons of God. (...) she kept the hope of being freed from the slavery of degradation, to know the freedom of glory given to the children of God. As we well know, the whole creation groans, it passes through the pains of a birth that lasts even longer. (...). For we have been saved, but it is in hope - St-Paul, Romans 8: 19-24
"Choose life..." CO2 housing -100% by 40% bullshit jobs How can we avoid 2021 = 2020? 2050? +x% unemployed/ covid Deep excluding poverty 3500 dead/ yr, 80% human "stupidity" Sustainable agriculture #1 - BENEFIT: Algorithms are a benefit for the co-creator and responsible man who discerns the good uses - Manifesto Espérance & Algorithmes
Decision s u p p o r t , a u t o m a t a , AI ... persona (humana) ex - machina From the DSS (inc. (Algorithm ic) Decis ion assisted driving): human Support - DSS elementary choice (inc. a Decision in the full sense: reason + free will few clicks) (Aristotle), therefore the prerogative of man ..... Autom aton (actuator) alone, neither of the beast nor of the machine ... Who decides, who's responsible? AI Robot Designer (or his Employer, his s/t ...) (interactive) Seter up to the learning agent: Data purveyor non-human elementary certifier choice, ... but its setting Consultant the rest Rational end user agent (learner) ... cf RGPD,Europ. Parliament/high risk
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / What are the benefits? Who decides the good (for others?)? What compromises? Vocation …. cf. Common Good 2 Efficient / Reliable / this s et of s ocial conditions that enable both groups and Robust / Risks their m em bers to achieve their perfection in a m ore total and eas y way. - Soc. Doc. Church 164-170 Systemic feedback 3 Explainable - Shared Control - Responsibility - 4 Autonomy - consent/ Confidence Freedom Of Conscience - Competition 7 Sobriety - contemplation 6 Human encounter – Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / Vocation 2 Efficient / Reliable / Robust / Risks Systemic feedback 3 Explainable - Shared Control - Responsibility - 4 Autonomy - consent/ Confidence Freedom Of Conscience - Competition 7 Sobriety - contemplation 6 Human encounter – Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
96% a cc. t o o u r M a ch in e Le a r n in g ... – t r u e ? Dis cern the red of blue over pa s t da ta to gues s the color of a future ca s e Feature# 2 Feature# 32 Feature# 1 Gizem Yetiş, Oza n Yetkin - A Novel Approa ch for Cla s s ifica tion of Structura l Elem ents in a 3D Model by Supervis ed Lea rning Septem ber 2018 Conference: Com puting for a better tom orrow - Proceedings of the 36th eCAADe Conference Volum e: 1 Feature# 25
Trustworthy ML models ? Dis cern the red of blue on pa s t da ta to gues s the color of a future ca s e ... Gizem Yetiş, Ozan Yetkin- A Novel Approa ch for Cla s s ifica tion of Structura l Elem ents in a 3D Model by Supervis ed Lea rning Septem ber 2018 Conference: Com puting for a better tom orrow - Proceedings of the 36th eCAADe Conference Volum e: 1 How does this dis cernm ent evolve a ccording to: ina ccura cy of the dots ? m ultiplica tion of da ta ? ... The blue-red interfa ce: how thick, wha t decis ion, wha t ris k?
Paradigm of decision support via model/algo /AI or human expertise Xj Observables j-ith past state of the world Yj Decisional event j data ... ... Science (ph. laws + init. cond.) Machine learning- IA Min Distance {Data ; Model (Xj,Yj)} m odel Observables: present state of the world diagnosis projected / Decision reco. Hum a n expert predictability = regula rity + obs erva bility
Inside generalisability error A good model : falsifiable & well understood failures regular determinist x -> y : same causes > same effects (in the future, as in Permanence of physical /phenomenological laws, rules ... ity the past) Stationnarity F(X; Y|x) remains stable: causes > effects at comparable Permanence of profiles - user behaviors etc. (static) frequencies - stable frequency of causes (in the future, as in the past) Stationnarity The cause-effect relationship F (X)t; Y|x,t) drifts according Climate change rising temperatures (dynamic) to a stable trend in frequency (hence learnable ...) Pb of human or environmental rapid feedback systems Interpolability A causes not observed but "close" to observed causes Simulation/Forward-Looking Requirements correspond to "close" effects Limitations of chaotic systems observ Fundamental & The state (and its variations): observable non- Scale >> Heisenberg uncertainty ability destructive Pb of economic forecasts technological Key observables are measurable with sufficient Difficult for climate, underground etc. precision/sampling via sensors economical Key observables are measurable at a reasonable Difficult for the human body, cultural preferences, ... cost/precision
Predictability, performance , u n ce r t a in t y ... good news for free will ! fingerprint Single outcome Low uncertainty Accounting automaton ST weather forecast, GPS router Real estate credit Auto translation score Cultural MT weather recommendation forecast Single outc. Multiple outc. uncertain uncertain Driving car Breast cancer screening
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / Vocation 2 Efficient / Reliable / Robust / Risks Systemic feedback 3 Explainable - Shared Control - Responsibility - 4 Autonomy - consent/ Confidence Freedom Of Conscience - Competition Increase Confidence in Decision transparency? 7 Sobriety - Explicable process?Readablemarginal sensitivity? Complete decision-making rules? contemplation 6 Human encounter Explicability - n/p reliability–? Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / Comprehensive consent ... including preferences, bias, risk, content filtering? Vocation Autonomy, including on the medical emergency? 2 Sacrifice / common Efficient good:/ eg. sharing health data, vaccination... / Reliable Freedom - competition, algo-monopolies Robust / Risks Systemic feedback 3 Explainable - Shared Control - Responsibility - 4 Autonomy - consent/ Confidence Freedom Of Conscience - Competition 7 Sobriety - contemplation 6 Human encounter – Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / Vocation 2 Efficient / Reliable / Robust / Risks Systemic feedback 3 Explainable - Shared Sans biais 4 indésirable, Autonomy ou-discriminant consent/ … selon quelles Control - Responsibilitypopulations - ?Freedom Quelle justice ? Qui en décide Confidence Of Conscience - ? Competition Cf. Option préférentielle pour le pauvre 7 Sobriety - contemplation 6 Human encounter – Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / Vocation 2 Efficient / Reliable / Robust / Risks Systemic feedback Releasing repetitive tasks to serve better (even if you lose some skills?) Increase skills by expanding experience/possible/cultural diversity 3 ExplainableDon't - Shared say it all... or not rightaway Control - Responsibility - 4 Autonomy - consent/ Confidence Freedom Of Conscience - Competition 7 Sobriety - contemplation 6 Human encounter – Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
A necessary bet ... Is it fa ir ? 1 Benefit / Garden / Vocation 2 Efficient / Reliable / Robust / Risks Systemic feedback Sober in automation i.e. elevating use for both the expert and the end user Without 3 deceptive anthropomorphism, Explainable - Shared virtual addiction or quantitative totalitarianism,... "Man doesn't live only Control on scores" - Responsibility - 4 Autonomy - consent/ Confidence Freedom Of Conscience - Competition 7 Sobriety - contemplation 6 Human encounter – Future Justice/discrimination - 5 Anthropomorphism of work - maturation - subsidiarity - participation deliberation
... theoretical expectations that are neither a lw a ys fe a s ib le n o r n e ce s s a r ily d e s ir a b le ...
The Re a l w o r ld - Et h ica l Te n s io n s in De cis io n Su p p o r t Performance rate - ST accuracy Explainability Personalization/contractual Non-discrimination /commutative equity justice Comfort - efficiency - ST value Personal achievement - MT value Performance intimacy Common good (incl third parties) Consent/autonomy Efficiency (rapidity - massivity) Security (ST and MT) efficiency... freedom...
Decision - m a k in g n a t u r e s a n d DSS s t a k e s Modera te ris k Undeterm ined ris k High ris k weather forecas t Recommendation cultural content Driving autonom ous vehicle Travel recom m endation Documentary research off-s ite clean Advertis ing dis play Recommendation meeting / CV Military robots Autom ation of s upport Real estate credit Facial recognition s ervices Weather forecast / performance / Medical em ergency diagnos is Chatbot/ com m ercial call sales / geotech ... ... centre Thermal comfort steering ... Fraud detection Non-urgent medical diagnosis .... Open exploration - forecast in law Multiple reco. - point forecast Autonomous agent – single recommendation
What ethics s h o u ld u lt im a t e ly discern t h e co m m o n g o o d Discerning via “smart” Predictive analys is / proba interactions Man -Man calculation. / learning on a supported by algo - expected utility function prudence Utilitarian ethics Decision Virtue ethics framework supported by algos Verification/ certification/ s af ety proces s by determ inis tic calculations Deontological ethics
Employment and AI Risks Short-term job cuts, or even jobs, through automation Medium-term cancellations of Luddites/competitive delays Dehumanizing work through excessive digital mediation Destroying skills by over-delegation to algorithms Increase career/exclusion silos Promises: Fluidizing the labour market Enlarge recruitment basis/bias of experience silos, atypical career brakes, long-term exclusion of people Deepen people's vocation and therefore joy at work Raising work by automating "stunning" tasks
Employment and AI "Creative destruction" is not necessarily as quick as announced 47% endangered jobs(2013) => +10%jobs (2020 pre-covid) It is those who seize transformations rather than undergo or repel them that shape the best opportunities
Creative Benefit Human Encounter co m p r o m is e s t o b e Shared Mastery Work fo u n d Garden Freedom of Consciousness (…) Vocation - t h e Subsidiarity Contemplation in im it a b le cr e a t ivit y o f t h e h u m a n p e r s o n is Vocation Participation n e ce s s a r y t o p u t Human Person Competition a lg o r it h m s a t t h e s e r vice Risks Universal Destination o f t h e co m m o n g o o d (…)
Annexes
Déterministe, probabiliste, prédictible, in ce r t a in , r is q u é ... Déterminé / déterministe / causal Indéterminé / probabiliste / ouvert Système sous-jacent observable - prédictible Système indéterminé - observabilité partielle - Dynamique incertaine Algo en phase prédictive Aléa données/processus d’apprentissage Prédiction randomisée Interactions / rétroactions MT Certification déterministe: ... un leurre ? Certification probabiliste: jamais de garantie (complexité des systèmes, biais sélection ...) certaine/ni acceptable a posteriori Biais cognitifs humains Liberté de l’agir humain / ouverture Expérience bornée du meilleur expert Perception humaine étendue
Manifeste Bienfait Rencontre Humaine Co m m e n t e n t r e p r e n d r e à Maîtrise Partagée Travail l’è r e d e s a lg o r it h m e s p o u r s e r vir le b ie n Jardin Liberté de Conscience co m m u n s a n s a s s e r vir Subsidiarité Contemplation Vocation Participation Personne Concurrence Risques Destination Universelle
OAD - b ie n d é t e r m in é , u n ivo q u e , p e u in ce r t a in En théorie: x(t°) -> D=y(t°) est univoque. X(t°) Données observées / percepts Signes, données, résultats de mesures au temps observé Pour bien décider, il suffit d’avoir Diagnostic / action réglée > les données x(t°) > Et l’apprentissage de la règle x->D D = Y(X(t°)) Estimation de l’état réel sous -jacent & Un algo peut mieux discerner que le meilleur expert décision univoque selon état humain Comptabiliser un mouvement - déclencher l’antispam - reconnaître une empreinte digitale ... Discernement pratique Précision & robustesse (généralisabilité, représentativité, stabilité num., observabilité, stationnarité, …) Contrôlabilité des erreurs et équité de traitement Développement des compétences / employabilité Taux de Perform ance - précis ion Explicabilité CT Confort - efficacité - valeur CT Réalis ation pers onnelle - valeur MT
OAD - b ie n d é t e r m in é , u n ivo q u e m a is in ce r t a in X(t°) Données observées / percepts En réalité: X(t°) observe imparfaitement l’état réel x(t°) Signes, données, résultats de mesures au temps observé et/ou la règle x-> d= y(t°) est insuffisamment apprenable Diagnostic faute d’assez de données fiables et/ou la règle x-> d est Y(t°) Estimation de l’état réel sous -jacent (au ambigüe, … et la décision porte sur un état futur y(tf) regard de la décision à prendre) dépendant d’une dynamique au pronostic non univoque Pronostic / évaluation La personne a-t-elle telle maladie ? S’agit-il d’un mouvement selon y(t°) frauduleux ? La personne est-elle coupable de tel acte ? ... d’impact Y(tf , d); L(Y(.)) Projection de l’état futur vraisemblable Discernement pratique et des conséquence (selon la décision) Périmètre d’observabilité machine vs. par l’agent humain Décision contextuelle Selon le traitement/la décision, que deviendra -t-elle (et son environnement ?) ? Contrôlabilité des erreurs et équité de traitement Temps de la décision et espace relation humaine/ de d(t°) Décision prise maturation Personnalisation / équité Non-discrimination / justice contractuelle commutative Précision Intimité Efficacité (rapidité & massivité) Consentement / autonomie
OAD - b ie n m u lt ivo q u e En réalité: x(t°) -> D=y(t°) est multivoque, soit par X(t°) Données observées / percepts insuffisante observabilité du réel soit par dilemmes de Signes, données, résultats de mesures au temps observé préférence. Pour bien décider, il faut Diagnostic / action réglée > une appréciation fidèle des possibles Yi(t°) > un modèle accepté de préférence aux conséquences Di D1 = Y1(X(t°)) Estimation de l’état réel sous -jacent & Un outil peut permettre à une personne de mieux décider D2 = Y2(X(t°)) ... décision multivoques Recommandation culturelle / rencontres, moteur de traduction automatique, prévision météo à 3 jours, ... Discernement pratique Consentement au modèle de préférences Ajustement des préférences selon l’horizon / asservissement aux silos / rétroaction sur l’apprentissage Elargissement vs. appauvrissement de l’expérience décisionnelle Confort - efficacité - valeur CT Réalis ation pers onnelle - valeur MT Précis ion Intim ité
Types d’algorithmes et enjeux d’OAD Enjeu fort ??? Enjeu faible Contrôle/commande de Exploration de données Exploration de données systèmes automatisés “fermée” (exfiltrage, réduction “ouverte” irréversible …) Recommandation (ou score, Prévision explicite en traduction/résumé, …) mono- Recommandation (ou score, incertitude résultat / classif. binaire traduction/résumé, …) multi- résultat ... Sélection automatisée ... Prévision de tendance simple ….
Prédictibilité, incertitudes, performance ... Univoque/ peu Prévision météo CT, routeur GPS Prédictibilité forte à CT, peu d’attente explicabilité perte de incertain compétences Annotation comptable Multivocité en réalité indésirable Reconnaissance empreinte ... Univoque / incertain Crédit immobilier Observ. Moyenne, dynamique incertaine Prévision météo MT Faible observ, dynamique incertaine Diag géotechnique Faible observabilité, dynamique certaine Diag dépistage cancer sein Faible observ, dynamique incertaine Multivoque / Recommandation culturelle, rencontres, recrutement ... Faible observabilité, Rétroaction forte, incertain Drug discovery Observabilité forte, dynamique incertaine Traduction automatique, résumé, analyse autom. CV, ... Observabilité moyenne, préférences incertaines Arrêt automatique véhicule autonome / dilemme tramway Observabilité moyenne (mais > l’humaine ?), préférences incertaines, dynamique très incertaine, rétrooaction comportementale forte
Libération ou asservissement ? Espionnage ou Suppression ou transformation Alliance ? des emplois ?
Machines numériques universelles Vin, parking/ expérience client ◆Ventes ◆Innovation Fabrication / conception produits ◆Production Coopératives, services à la personne, recrutement BTP, flottesChain ◆Supply véhicules Compta, droit,/services ◆Contrôle back auxoffice DATA entreprises Modèle = industrialise la décision ◆... Algorithme = démultiplie l’exécution Int. Artificielle = rend apprenante l’interaction
You can also read