Data Science and Decisions 2022: Week 4

Transcript

1. DATA SCIENCE AND DECISION MAKING Machines and Decisions Will Lowe Hertie School Data Science Lab 2022-03-09
2. MACHINE LEARNING AND OTHER ADM 1 Machines, automation, and decisions Two converging case studies: deciding and knitting Resistance to automation Stakeholders: science, government and society ree ways to think about automated decision making
3. MACHINE LEARNING AND OTHER ADM 2 We’ll take a broad de nition of machine learning for decision making → any kind of algorithmically de ned, automated process ending with an action Examples that di er only in implementation → Rule-based expert system, neural network, y-by-wire piloting system We’ll focus on issues around automation Claim (Algorithm Watch Report ): we’re no longer just automating society. We have automated it already Claim : → e problems of automation have very little to do with machines, or learning
4. CONVERGING CASE STUDIES 3 Machine learning Machine weaving
5. MACHINE WEAVING 4 I → pre- : designs are realized by a ‘draw-boy’ moving threads, directed by a weaver → : Basile Bouchon and Jean-Baptiste Falcon gured out how to realize a design using punched cards → : Jacques de Vaucanson arranged the a punched tape to move to the next row, a ‘carriage return’ (a weaver provided the power source) → Joseph Marie Jacquard invents and popularizes the programmable loom A portrait in , cards
6. MACHINE WEAVING 5 Powered looms were possible since Cartwright and perfected by in the ‘Lancashire Loom’ Steady removal of the human from designing, realizing, and powering the process
7. MACHINE WEAVING 5 Powered looms were possible since Cartwright and perfected by in the ‘Lancashire Loom’ Steady removal of the human from designing, realizing, and powering the process At the same time → - e Continental System (a trade blockade from most of Europe) → Di culties with American trade Weaving gets with the program
8. RESISTANCE 6 ‘Luddites’ destroying a Jacquard loom s: ‘Luddites’ destroyed weaving machines, rioted, and assassinated a mill owner What was this resistance to? → Machines themselves? → British macroeconomic policy? → Lack of social safety net? → Worsening factory working conditions? → Removal / denigration of the intrinsic value of labour
9. RESISTANCE 6 ‘Luddites’ destroying a Jacquard loom s: ‘Luddites’ destroyed weaving machines, rioted, and assassinated a mill owner What was this resistance to? → Machines themselves? → British macroeconomic policy? → Lack of social safety net? → Worsening factory working conditions? → Removal / denigration of the intrinsic value of labour : ‘Lancaster Loom’ invented : Friedrich Engels is sent to Salford (Greater Manchester) to oversee his father’s weaving factory
10. PREFERENCE AGGREGATION... 7 Suppose one of these men, as I have seen them, – meagre with famine, sullen with de- spair, careless of a life which your lordships are perhaps about to value at something less than the price of a stocking-frame – suppose this man surrounded by the chil- dren for whom he is unable to procure bread at the hazard of his existence, about to be torn for ever from a family ... Are we aware of our obligations to a mob? It is the mob that labour in your elds, and serve in your houses – that man your navy, and recruit your army – that have en- abled you to defy all the world, – and can also defy you, when neglect and calamity have driven them to despair. You may call the people a mob, but do not forget that a mob too o en speaks the sentiments of the people. Lord Byron, Feb th, on the Frame Work Bill (Hansard link)
11. ...INSIDE A UNITARY STATE 8 e Frame Work Bill passed and turned into the Destruction of Stocking Frames, etc. Act ( Geo c. ) Made destruction of looms a ‘capital felony’ → Previously - years in a penal colony(!) → Later changed to transportation → And then back to a capital punishment e Luddite movement was e ectively suppressed: many hanged, killed by troops, imprisoned... In the meantime, Byron’s daughter Ada was becoming a programmer Ada Lovelace
12. ENGINES 9 Charles Babbage and Ada Lovelace (nee Byron) worked on a mathematical engine for → Logarithms → Navigation calculations Quickly surpassed by its programmable successor e Analytical Engine → Unfortunately, before the Di erence Engine was completed e Analytical Engine was → Programmable → Turing Complete e Di erence Engine (∼ )
13. THE ANALYTICAL ENGINE 10 e Analytical Engine (∼ ) A program
14. WHAT DOES GOVERNMENT WANT? 11 e British government funded most of the development Why? Because it wanted the output: → Astronomical and mathematical tables When the machines were not built or abandoned they were (perhaps understandably) unhappy... → Babbage could not or did not see this dynamic → Lovelace didn’t need to → ey were building something that ‘did’ mathematics Science!
15. WHAT DOES GOVERNMENT WANT? 11 e British government funded most of the development Why? Because it wanted the output: → Astronomical and mathematical tables When the machines were not built or abandoned they were (perhaps understandably) unhappy... → Babbage could not or did not see this dynamic → Lovelace didn’t need to → ey were building something that ‘did’ mathematics Science! years later → Turing and Welchman knew – and built the ‘Bombe’
16. WHAT DOES HUMANITY WANT? 12 People have been torn about the emancipatory and/or dehumanizing possibilities ever since As soon as labour in the direct form has ceased to be the great well-spring of wealth, labour time ceases and must cease to be its measure, and hence exchange value [must cease to be the measure] of use value. e surplus labour of the mass has ceased to be the condition for the development of general wealth, just as the non-labour of the few, for the development of the general powers of the human head. With that, production based on exchange value breaks down, and the direct, material production process is stripped of the form of penury and antithesis. e free development of individualities[, ...] the general reduction of the necessary labour of society to a minimum, which then corresponds to the artistic, scienti c etc. development of the individuals in the time set free, and with the means created, for all of them. Marx Grundrisse ch. para. (the ‘Fragment on Machines’)
17. HOW TO THINK ABOUT ADM 13 T → Rhetorical frames → Functional roles → Normative frameworks
18. RHETORICAL FRAMES 14 E Cameron ( ) ‘Aliens’ Augmenting human decision making: ‘harder, better, faster, stronger’ A Leckie ( ) Either a ‘new’ decision maker, or a form of collective decision making Fall ( ) G Rabbi Judah Loew ben Bezalel ( s) ‘Josef’ c.f. McElreath ( ) Statistical Rethinking (ch. )
19. HOW TO THINK ABOUT ADM 15 Google glass (it was a thing) F Locating humans in decision making → Human decisions about action, individual and unaided → Human decisions about action, in institutions → Human decisions about action, aided by calculation tools → Human decisions about inputs to calculation e.g. probability or loss elicitation, as trigger, and as oversight or governance → Human decision outcomes (behaviour) as input, e.g. collaborative ltering → Combinations of the elements above, e.g. automated trading → Human decisions about goals/losses, e.g. self-driving cars, automated fraud checking
20. NORMATIVE FRAMEWORKS 16 C → Judge a decision technology by its outputs e.g. utility → O en so ened to ‘rule consequentialism’, e.g. ‘maximize expected utility’ Implications → Judge the fairness of a decision technology by its actual and its counterfactual outputs → Transparency is not particular important
21. NORMATIVE FRAMEWORKS 16 C → Judge a decision technology by its outputs e.g. utility → O en so ened to ‘rule consequentialism’, e.g. ‘maximize expected utility’ Implications → Judge the fairness of a decision technology by its actual and its counterfactual outputs → Transparency is not particular important D → Judge a technology by its operating principles → Most naturally applied to rule-based decision systems Implications → Judge the fairness of a technology is determined by the operating principles it realizes → Transparency is required
22. TRANSPARENCY? 17 Transparency is o en considered a virtue in decision making, but identifying it is tricky → Not quite visibility. Seeing the mechanism may not help me much → Not quite explanation. You can try to explain EU anti-trust law to me but I’ll fall asleep → Not quite explanability. What counts as an explanation is contextually variable → Not quite justi cation: e rule you show me may not be the cause, e.g. ‘parallel construction’, e.g. Enigma → Not quite manipulability. Seeing and doing give di erent information We might like these properties for themselves Local Interpretable Model-agnostic Explanation (Ribeiro et al., )
23. A REPRESENTATIVE CONCERN 18 Algorithms are neither “neutral” nor “objective” even though we tend to think that they are. ey replicate the assumptions and beliefs of those who decide to deploy them and program them. Humans, therefore, are, or should be, responsible for both good and bad algorithmic choices, not “algorithms” or ADM systems. e machine may be scary, but the ghost within it is always human. And humans are complicated, even more so than algorithms. Algorithm Watch ( ) What do we make of this with respect to functional roles, rhetorical frames, and normative assumptions?
24. MACHINE LEARNING AND OTHER ADM 19 Machines, automation, and decisions Two converging case studies: deciding and knitting Resistance to automation Stakeholders: science, government and society ree ways to think about automated decision making
25. REFERENCES 20 Algorithm Watch. ( , October). Automating society . Algorithm Watch and Bertelsmann Sti ung. URL. Fall, I. ( ). I sexually identify as an attack helicopter [magazine]. Clarkesworld Magazine, ( ). URL. Leckie, A. ( ). Ancillary justice. Orbit. Marx, K. ( ). Grundrisse (M. Nicolaus, Ed.). Penguin. (Original work published ) McElreath, R. ( ). Statistical Rethinking: A Bayesian Course with Examples in R and Stan. CRC Press, Taylor & Francis Group. Ribeiro, M. T., Singh, S., & Guestrin, C. ( , August ). Why should i trust you?’: Explaining the predictions of any classi er. arXiv: . [cs, stat]. Retrieved November , , from URL.