Part V: Considerations and conclusions

In this part, we first include a set of written contributions by other scholars that were not presented at the workshop but provided relevant input for our final considerations.

Then, we provide a set of conclusions to the workshop and directions for future work in the HUMAINT project.

20 Characterising the trajectories of artificial and natural intelligence

20.1 State of the art: An Atlas of Intelligence

The comparison between artificial and human intelligence is usually done in an informal and subjective way, often leading to contradicting assessments (Kirsh 1991, Hayles 1996, Brooks 1997, Pfeifer 2001, Shah et al 2016, Lake et al. 2017, Tegmark 2017, Marcus 2018). This is especially problematic because of the pace of the epistemological change. Our understanding of intelligence is rapidly progressing from new discoveries in comparative cognition, neuroscience and artificial intelligence. However, there is a possibly more relevant ontological change: artificial intelligence is creating new kinds of systems, and it is hence extending the landscape of intelligence (Sloman, 1984). Moreover, it is still not fully recognised —and certainly not well understood— that these technological changes are also affecting human cognition. Put it simply: because of AI, humans now think differently. Overall, and by all means, we have a moving target problem. Can we anticipate these trajectories?

In the first place, we need ways of assessing what AI systems can do, what they will be able to do in the near future, and the pathways and resources that will be needed to get there. Indeed, we need a common framework to determine which kinds of AI or hybrid systems in this landscape of intelligence are even desirable (needed for society) or possibly undesirable (dangerous, too similar to some profession profiles, animals or humans, etc.). For a recent symposium about this see: http://kindsofintelligence.org/.

The discussion must not be limited to the way society is affected: the irruption of AI systems with new capabilities may trigger a range of alterations in the very way human cognition works. The changes in memory capabilities, development trajectories and learning patterns that we are already observing because of the use of technology (negative Flynn effect, Google effect, etc.) can distinctly be regarded as cognitive atrophy or enhancement, but are about to change our psychometric profiles in possibly radical ways.

The Leverhulme Centre for the Future of Intelligence (http://lcfi.ac.uk) is working on a new initiative, an atlas of intelligence, to cover a relevant portion of the past, present and future landscape of intelligence: including humans, non-human animals, AI systems, hybrids and collectives thereof.

The atlas will be based on a set of dimensions, either behavioural features (i.e., the functionalities, cognitive abilities and personality traits) or physical features (i.e., the mechanisms, kinds of sensors and actuators, body morphology, computational or neurological resources). The atlas will allow users to make several projections and aggregations to a smaller number of dimensions. Also, despite the framework not being hierarchical, once populated, it could be converted into different kinds of taxonomies by using different distance/similarity metrics (as homology or analogy have been used for living systems) and also exploring a continuum from specialised (task-specific) systems to more general (task-independent) systems, including a developmental perspective.

The initiative is at an early stage and we welcome associates and contributors. More information about the specification and prospective maps to be considered for the atlas can be found in (Bhatnagar et al. 2017, 2018).

20.2 Challenges

• How can we characterise current and future AI systems in terms of cognitive abilities (Hernández-Orallo 2017a,b) and compare them to humans? o Are the new evaluation platforms (Castelvecchi 2016, Hernández-Orallo et al.

2017. going in the right direction? How do the AI milestones relate or compare to the milestones in child development or animal evolution? o Can we develop an ability-oriented analysis of job automation rather than task-oriented (Frey and Osborne 2017, Brynjolfsson and Mitchell 2017)?

• How can we characterise the changes in human cognition originating from technology and, most especially, from the interaction, replacement or enhancement with AI systems? o How can we analyse the locations and trajectories of human intelligence and AI progress? o How can the new “cognitive ecosystems” (Hutchins 2010), including humans and machines, be affected by these future changes of intelligence and their effect on dominance topologies (Cave 2017, de Weerd et al. 2017)?

References

— Bhatnagar, S., Alexandrova, A., Avin, S., Cave, S., Cheke, L., Crosby, M., Feyereisl,

J., Halina, M., Loe, B.S., O hEigeartaigh, S., Martínez-Plumed, F., Price, H., Shevlin, H., Weller, A., Winfield, A. and Hernández-Orallo, J. “A first survey on an atlas of intelligence”. (2017) http://users.dsic.upv.es/~flip/papers/Bhatnagar18_SurveyAtlas.pdf — Bhatnagar S., Alexandrova A., Avin S, Cave S, Cheke L, Crosby M, Feyereisl J, Halina

M., Loe B.S., O hEigeartaigh S., Martínez-Plumed F., Price H., Shevlin H., Weller A., Winfield A. and Hernández-Orallo, J. (2018) “Mapping Intelligence: Requirements and Possibilities”, in Muller V. (ed) Philosophy and Theory of Artificial Intelligence. — Brooks, R. A. (1997). "From earwigs to humans." Robotics and autonomous systems

20.2-4, 291-304. — Brynjolfsson, E. and Mitchell, T. (2017). “What can machine learning do? Workforce implications”, Science 22 Dec 2017: Vol. 358, Issue 6370, pp. 1530-1534, DOI:

10.1126/science.aap8062, http://science.sciencemag.org/content/358/6370/1530 — Castelvecchi, D. (2016). “Tech giants open virtual worlds to bevy of AI programs”. Nature Vol 540, Issue 7633, pp. 323-324, https://www.nature.com/news/tech-giants- open-virtual-worlds-to-bevy-of-ai-programs-1.21151.

— Cave, S. (2017) “On the dark history of intelligence as domination” AEON, 21 February, 2017.

— de Weerd, H., Verbrugge, R. and Verheij, B. (2017) “Negotiating with other minds: the role of recursive theory of mind in negotiation with incomplete information” Autonomous Agents and Multi-Agent Systems, March 2017, Volume 31, Issue 2, pp 250–287.

— Frey, C. B., and Osborne, M. A. (2017). “The future of employment: How susceptible are jobs to computerisation?” Technological Forecasting and Social Change, 114, 254–280, 2017.

— Hayles, N. K. Narratives of artificial life. Future Natural: Nature, Science, Culture. (1996). London: Routledge, 146-64.

— Hernández-Orallo, J. (2017). The Measure of All Minds: Evaluating Natural and Artificial Intelligence, Cambridge University Press.

— Hernández-Orallo, J. (2017). "Evaluation in artificial intelligence: from task-oriented to ability-oriented measurement", Artificial Intelligence Review 48 (3), 397-447, 2017

— Hernández-Orallo, J., Baroni, M., Bieger, J., Chmait, N., Dowe, D.L., Hofmann, K., Martínez-Plumed, F., Strannegård, C. and Thórisson, K.R. (2017). "A New AI Evaluation Cosmos: Ready to Play the Game?" AI Magazine, Association for the Advancement of Artificial Intelligence.

— Hutchins, E. (2010). "Cognitive ecology" Topics in cognitive science 2.4 (2010): 705-

— Kirsh, D. (1991). "Today the earwig, tomorrow man?" Artificial intelligence 47.1-3: 161-184.

— Lake, B.M., Ullman, T.D., Tenenbaum, J.B., Gershman S.J.. “Building machines that learn and think like people”. Behavioral and Brain Sciences, 2017.

— Marcus G. “Deep Learning: A Critical Appraisal” arXiv preprint arXiv:1801.00631. 2018 (also https://medium.com/@GaryMarcus/in-defense-of-skepticism-about-deep-learning- 6e8bfd5ae0f1)

— Pfeifer, R.. "Embodied artificial intelligence 10 years back, 10 years forward." Informatics. Springer, Berlin, Heidelberg, 2001.

— Shah, H., Warwick, K., Vallverdú, J., & Wu, D.” Can machines talk? Comparison of Eliza with modern dialogue systems”. Computers in Human Behavior, 58, 278-295,

— Sloman, A. The Structure and Space of Possible Minds. School of Cognitive Sciences, University of Sussex, in The Mind and the Machine: philosophical aspects of Artificial Intelligence, ed. Stephen Torrance, Ellis Horwood, pp 35-42,1984.

— Tegmark, M. Life 3.0: Being Human in the Age of Artificial Intelligence. Knopf, 2017.

21 Considerations related to cognitive development children

Nuria Sebastian

Universitat Pompeu Fabra

My comments will turn around two issues, both related to the concept of “development” that complements the current views.

21.1 Differences in machinery

There is a fundamental difference between “Artificial Intelligence” and “Human Intelligence” related to the enormous changes that the Human hardware (the brain) undergoes during life. Comparatively, artificial hardware undergoes relatively small changes (and the changes are not just in its size, but also in qualitative aspects).

The architecture of the human brain changes in fundamental dimensions during life. The most dramatic changes take place during the first years of life (I will not refer to prenatal changes / learning, because the point I want to make does not require to address this period, but there is learning during this time). The number of neurons increases exponentially mostly prenatally, but the number of synapses changes in a complex way after birth. It is worth noticing that changes do not take place in an uniform way across the brain. In general, sensory-related areas develop very quickly (reaching adult levels by the end of the first year of life), while “thinking-planning” areas (frontal areas) reach adult levels after puberty. On top of complex changes in the number of synapses, the amount of myelination (related to the effectiveness of neural transmission) diminishes with age (though it does not disappear).

These specific patterns impose important constraints the way the brain processes the inputs it receives. An important feature of brain development is the extraordinary synchronization between maturation of different brain areas: they become functional when they are needed. For instance: association areas become functional when “lower” areas are effective: there is no energy waste by having areas “waiting” for appropriate inputs. Functionality is the product of an exquisite interplay between internal development (gene-regulated, more prevalent early in life) and external input (more prevalent late in life).

On top of these substantial hardware changes, there are other important development-related changes in neurotransmitters and hormones that will have dramatic consequences on the way the brain functions across life. A clear case is the changes in sleep patterns taking place in life. Newborns spend most of their time sleeping, while elderly people tend to sleep very few hours. There are fundamental changes in the way the brain functions during sleep (not only quantitative, but also qualitatively) and it is well-known the critical importance of sleep in memory-consolidation.

Finally, newborns (from 4 hours to 4 days of age) and very young infants are able to perform complex computation over different types of signals (that seem to be relatively experience-independent). For instance, newborns can notice the difference between some (human) languages, such as Dutch and Japanese, if played forwards, but not backwards (as other species such as cotton-top tamarin monkeys and long evans rats). It is not until 5 months of age that human infants can distinguish English from Dutch, or Spanish from Catalan (it is worth noticing that by six months, infants already know several words). In a different domain, newborns prefer to orient to stimuli with a human face configuration than to a random one (Morton and Johnson, 1991).

In summary, there are essential differences between human and artificial “hardware” and they entail critical specificities regarding human learning.

21.2 Developmental changes in interaction with computers

The presentations have assumed that users interacting with AI systems are adults. However, there are important differences in the way children and adults deal with artificial systems, and this is an under-studied field.

One of the few existing studies has investigated the well-known Uncanny Valley effect. There is a vast literature investigating the fact that (human) adults feel uncomfortable when interacting with very human-like avatars/robots.

Figure 5. Uncanny Valley effect.

There is evidence indicating that such effect may be acquired. Young children (under 9 years) do not find “creepy” such very human-like avatars, importantly the feeling of “creepiness” is related to children’s assumption that such avatars have human-like minds (Brink, Gray and Wellman, 2017).

The investigation of how children interact with machines is a virtually unexplored field. Such investigations are critical when considering the use of AI and robots in educational (and health) environments.

References

— Kimberly A., Brink, K. G. and Wellman, H. M. (2017). Creepiness Creeps In: Uncanny Valley Feelings Are Acquired in Childhood. Child development, https://doi.org/10.1111/cdev.12999

— Morton, J., and Johnsson, M. H. (1991). CONSPEC and CONLERN: a two-process theory of infant face recognition. Psychol Rev. 1991 Apr;98(2):164-81.

22 The tyranny of data? The bright and dark sides of algorithmic decision making for public policy making

Nuria Oliver

Vodafone Research and Datapop alliance

22.1 Statement: Data-driven Algorithms for Public Policy Making

Today's vast and unprecedented availability of large-scale human behavioral data is profoundly changing the world we live in. Massive streams of data are available to train algorithms which, combined with increased analytical and technical capabilities, are enabling researchers, companies, governments and other public sector actors to resort to data-driven machine learning-based algorithms to tackle complex problems (Gillespie,

2014). Many decisions with significant individual and societal implications previously made by humans alone --often by experts--are now made or assisted by algorithms, including hiring, lending (Khandani et al., 2010), policing (Wang et al., 2013), criminal sentencing (Barry-Jester et al., 2015), and stock trading. Data-driven algorithmic decision making may enhance overall government efficiency and public service delivery, by optimizing bureaucratic processes, providing real-time feedback and predicting outcomes (Sunstein, 2012). In a recent book with the evocative and provocative title Technocracy in America", international relations expert Parag Khanna argued that a data-driven direct technocracy is a superior alternative to today's (alleged) representative democracy, because it may dynamically capture the specific needs of the people while avoiding the distortions of elected representatives and corrupt middlemen (Khanna, 2017). Human decision making has often shown significant limitations and extreme bias in public policy, resulting in inefficient and/or unjust processes and outcomes (Fiske, 1998; Samuelson and Zeckhauser, 1998). The turn towards data-driven algorithms can be seen as a reflection of a demand for greater objectivity, evidence-based decision-making, and a better understanding of our individual and collective behaviors and needs. At the same time, scholars and activists have pointed to a range of social, ethical and legal issues associated with algorithmic decision-making, including bias and discrimination (Barocas and Selbst, 2016; Ramirez et al., 2016) and lack of transparency and accountability (Citron and Pasquale, 2014; Pasquale, 2015; Zarsky, 2016). For example, Barocas and Selbst (2016) showed that the use of algorithmic decision making processes could result in disproportionate adverse outcomes for disadvantaged groups, in ways suggestive of *discrimination*. Algorithmic decisions can reproduce and magnify patterns of discrimination, due to decision makers' prejudices or reflect the biases present in the society. A recent study by ProPublica of the COMPAS Recidivism Algorithm (an algorithm used to inform criminal sentencing decisions by predicting recidivism) found that the algorithm was significantly more likely to label black defendants than white defendants, despite similar overall rates of prediction accuracy between the two groups (Angwin et al., 2016). Along this line, a nominee for the National Book Award, Cathy O'Neil's book, Weapons of Math Destruction", details several case studies on harms and risks to public accountability associated with big data-driven algorithmic decision-making, particularly in the areas of criminal justice and education. In 2014, the White House released a report titled ``Big Data: Seizing opportunities, preserving values''¹⁰ highlighting the discriminatory potential of Big Data, including how it could undermine longstanding civil rights protections governing the use of personal information for credit, education, health, safety, employment, etc. For example, data-driven algorithmic decisions about applicants for jobs, schools or credit may be affected by hidden biases that tend to flag individuals from particular demographic groups as unfavorable for such opportunities. Such outcomes can be self- ¹⁰https://obamawhitehouse.archives.gov/sites/default/files/docs/20150204_Big_Data_Seizing_Opportunities_Pr eserving_Values_Memo.pdf

reinforcing, since systematically reducing individuals' access to credit, employment and education will worsen their situation, and play against them in future applications. For this reason, a subsequent White House report called for ``equal opportunity by design" as a guiding principle in those domains. Furthermore, the White House Office of Science and Technology Policy, in partnership with Microsoft Research and others, has co-hosted several public symposiums on the impacts and challenges of algorithms and Artificial Intelligence, specifically relating to social inequality, labor, healthcare and ethics¹¹

At the heart of the matter is the fact that technology outpaces policy in most cases; here, governance mechanisms of algorithms have not kept pace with technological development. Several researchers have recently argued that current control frameworks are not adequate for situations in which a potentially unfair or incorrect decision is made by a computer (Barocas and Selbst, 2016).

Fortunately, there is increasing awareness of the detrimental effects of discriminatory biases and opacity of some data-driven algorithmic decision-making systems, and of the need to reduce or eliminate them. A number of research and advocacy initiatives are worth noting, including the Data Transparency Lab¹², a ``community of technologists, researchers, policymakers and industry representatives working to advance online personal data transparency through research and design", and the DARPA Explainable Artificial Intelligence (XAI) project¹³. A tutorial on the subject was held at the 2016 ACM Knowledge and Data Discovery conference (Hajian et al., 2016). Researchers from New York University's Information Law Institute --such as Helen Nissenbaum and Solon Barocas-- and Microsoft Research --such as Kate Crawford and Tarleton Gillespie-- have held several workshops and conferences these past few years on the ethical and legal challenges related to algorithmic governance and decision-making¹⁴.

This chapter is a summary of the content discussed by Lepri et al. (2017a, 2017b), where the authors highlight the need for social good decision-making algorithms (i.e. algorithms strongly influencing decision-making and resource optimization of public goods, such as public health, safety, access to finance and fair employment) to provide transparency and accountability, to only use personal information --created, owned and controlled by individuals--with explicit consent, to ensure that privacy is preserved when data is analyzed in aggregated and anonymized form, and to be tested and evaluated in context by means of living lab approaches involving citizens.

The opportunity to significantly improve the processes leading to decisions that affect millions of lives is huge. As researchers and citizens, I believe that we should not miss on this opportunity. Hence, I would like to encourage the larger community --researchers, practitioners, policy makers--in a variety of fields --computer science, sociology, economics, ethics, law--to join forces so we can address today's limitations in data-driven decision-making and contribute to fairer and more transparent decisions with clear accountability, within an ethical framework and developed by diverse teams so they can achieve significant positive impact.

22.2 Challenges

There are several limitations and risks in the use of data-driven predictive models informing decisions that might impact the daily lives of millions of people. Namely:

21.2.1. Discrimination: Algorithmic discrimination may arise from different sources. First, input data into algorithmic decisions may be poorly weighted, leading to disparate impact. For example, as a form of indirect discrimination, overemphasis of zip code within predictive policing algorithms can lead to the association of low-income African-American neighborhoods with areas of crime and as a result, the application of specific 11 https://www.whitehouse.gov/blog/2016/05/03/preparing-future-artificial-intelligence http://www.datatransparencylab.org/ http://www.darpa.mil/program/explainable-artificial-intelligence http://www.law.nyu.edu/centers/ili/algorithmsconference

targeting based on group membership (Christin et al., 2015). Second, discrimination can occur from the decision to use an algorithm itself. Categorization can be considered as a form of direct discrimination, whereby algorithms are used for disparate treatment (Diakopoulos, 2015). Third, algorithms can lead to discrimination as a result of the misuse of certain models in different contexts (Calders and Zliobaite, 2013). Fourth, in a form of feedback loop, biased training data can be used both as evidence for the use of algorithms and as proof of their effectiveness (Calders and Zliobaite, 2013). The use of algorithmic data-driven decision processes may also result in individuals being denied opportunities based not on their own action but on the actions of others with whom they share some characteristics. For example, some credit card companies have lowered a customer's credit limit, not based on the customer's payment history, but rather based on analysis of other customers with a poor repayment history that had shopped at the same establishments where the customer had shopped (Ramirez et al., 2016). While several proposals have been made in the literature to tackle algorithmic discrimination and maximize fairness, we feel the urgency to establish a call for action bringing together researchers from different fields (including law, ethics, political philosophy and machine learning) to devise, evaluate and validate in the real-world alternative fairness metrics for different tasks. In addition to this empirical research, we believe it will be necessary to propose a modeling framework --supported by empirical evidence-- that would assist practitioners and policy makers in making decisions aided by algorithms that are maximally fair.

21.2.2. Lack of Transparency/Opacity: Transparency, which refers to the understandability of a specific model, can be a mechanism that facilitates accountability. More specifically, transparency can be considered at the level of the entire model, at the level of individual components (e.g. parameters), and at the level of a specific algorithm. In the strictest sense, a model is transparent if a person can contemplate the entire model at once. Thus, models should be characterized by low computational complexity. A second and less strict notion of transparency might be that each part of the model (e.g. each input, parameter, and computation) admits an intuitive explanation. A final notion of transparency might apply at the level of the algorithm, even without the ability to simulate an entire model or to intuit the meaning of its components. However, the ability to access and analyze behavioral data about customers and citizens on an unprecedented scale gives corporations and governments powerful means to reach and influence segments of the population through targeted marketing campaigns and social control strategies. In particular, we are witnessing an information and knowledge asymmetry situation where a powerful few have access and use resources and tools that the majority do not have access to, thus leading to an --or exacerbating the existing-- asymmetry of power between the state and big companies on one side and the people on the other side, conceptualized as a “digital divide" (Boyd and Crawford, 2012). In addition, the nature and use of various data-driven algorithms for social good, as well as the lack of computational or data literacy among citizens (Bhargava et al., 2015), makes algorithmic transparency difficult to generalize and accountability difficult to assess (Pasquale, 2015). Burrell (2016) has provided a useful framework to characterize three different types of opacity in algorithmic decision-making: (1) intentional opacity, whose objective is the protection of the intellectual property of the inventors of the algorithms. This type of opacity could be mitigated with legislation that would force decision-makers towards the use of open source systems. The new General Data Protection Regulations (GDPR) in the EU with a “right to an explanation" starting in May of 2018 is an example of such legislation. But powerful commercial and governmental interests will make it difficult to eliminate intentional opacity; (2) illiterate opacity, due to the fact that the vast majority of people lack the technical skills to understand the underpinnings of algorithms and machine learning models built from data. This kind of opacity might be attenuated with stronger education programs in computational thinking and “algorithmic literacy" and by enabling independent experts to advise those affected by algorithmic decision-making; and (3) intrinsic opacity, which arises by the nature of certain machine learning methods that are difficult to interpret (e.g. deep learning models). This opacity is well known in the machine learning community (usually referred to as the interpretability problem).

21.2.3. Computational violations of privacy: Reports and studies (Ramirez et al., 2016) have focused on the misuse of personal data disclosed by users and on the aggregation of data from different sources by entities playing as data brokers with direct implications in privacy. An often overlooked element is that the computational developments coupled with the availability of novel sources of behavioral data (e.g. social media data) now allow inferences about private information that may never have been disclosed. This element is essential to understand the issues raised by these algorithmic approaches, as has become apparent in the recent Facebook/Cambridge Analytica data scandal¹⁵. 21.2.4. Data Literacy: It is of paramount importance that we devote resources to computational and data literacy programs aimed at all citizens, from children to the elderly. Otherwise, it will be very difficult, if not impossible, for us collectively as a society to make informed decisions about technologies that are not fully understood (Bhargava et al., 2015). 21.2.5. Unclear accountability: As more decisions that affect the lives of thousands of people are automatically made by algorithms, we need clarity on who is responsible for the decisions made by them or with algorithmic support. Transparency is generally thought as a key enabler of accountability. However, transparency and auditing do not necessarily suffice for accountability. In fact, in a recent paper Kroll et al. (2017) have introduced computational methods able to provide accountability even when some information is kept hidden. 21.2.6. Lack of ethical frameworks: Data-driven algorithmic decision-making poses important ethical dilemmas regarding what would be an appropriate course of action to take based on the inferences carried out by the algorithms or on the specific situation that the algorithm is acting upon. Hence, practitioners, developers, researchers and policy makers who would use data-driven algorithms to support or automatically make decisions would need to ensure that such decisions are made in accordance with a pre-defined and commonly accepted ethical framework. There are several examples of ethical principles proposed in the literature for this purpose¹⁶¹⁷ and institutes and research centers, such as the Digital Ethics Lab in Oxford or the AI Now Institute at NYU. However, it is an open question how to properly incorporate ethical principles in data-driven algorithmic decision making processes in addition to ensuring that all the developers and professionals involved comply with a clear Code of Conduct and Ethics. 21.2.7. Lack of diversity: Given the broad set of use cases that data-driven algorithms might be apply to**,** it is important to reflect on the diversity of the teams that generated such algorithms. To date, the development of the state-of-the-art data-driven, machine learning-based algorithms has been carried out by somewhat homogeneous groups of computer scientists. Moving forward, we need to ensure that the teams are diverse both in terms of areas of expertise and demographics –particularly gender. https://en.wikipedia.org/wiki/Facebook%E2%80%93Cambridge_Analytica_data_scandal https://www.wired.com/story/should-data-scientists-adhere-to-a-hippocratic-oath/ https://futureoflife.org/ai-principles/

Figure 6. Summary of requirements for positive data-driven disruption.

While this is an exciting time for researchers and practitioners in this new field of computational social sciences, we need to be aware of the risks associated with these new approaches to decision making, including violation of privacy, lack of transparency and diversity, information and knowledge asymmetry, social exclusion and discrimination. I would like to highlight three human-centric requirements that we consider to be of paramount importance to enable positive disruption of data-driven policy-making: user-centric data ownership and management; algorithmic transparency and accountability; and living labs to experiment with data-driven policies in the wild. It will be only when we honor these requirements that we will be able to move from the feared tyranny of data and algorithms to a data-enabled model of democratic governance running against tyrants and autocrats, and for the people.

For the readers interested in the topic, they can find an extended version of this chapter in (Lepri et al., 2017a; Lepri et al., 2017b).

References

— Angwin, J., Larson, J., Mattu, S., Kirchner, L.: Machine bias. ProPublica (2016). URL https://www.propublica.org/article/machine-bias-risk-assessments-in-criminal-sentencing

— Barocas, S., Selbst, A.: Big data's disparate impact. California Law Review 104, 671{732 (2016)

— Barry-Jester, A.M., Casselman, B., Goldstein, D.: The new science of sentencing. The Marshall Project (2015). URL https://www.themarshallproject.org/2015/08/04/the-new-science-of- sentencing.

— Bhargava, R., Deahl, E., Letouze, E., Noonan, A., Sangokoya, D., Shoup, N.: Beyond data literacy: Reinventing community engagement and empowerment in the age of data. Data-Pop Alliance White Paper Series (2015). URL http://datapopalliance.org/wp-content/uploads/2015/11/Beyond-Data-Literacy-2015.pdf

— Boyd, d., Crawford, K.: Critical questions for big data: Provocations for a cultural, technological, and scholarly phenomenon. Information, Communication, & Society 15(5), 662{679 (2012)

— Burrell, J. (2016). How the machine `thinks': Understanding opacity in machine learning algorithms. Big Data & Society 3(1)

— Calders, T., Verwer, S.. (2010). Three naive bayes approaches for discrimination-free classification. Data Mining and Knowledge Discovery 21(2), 277:292.

— Calders, T., Zliobaite, I.: Why unbiased computational processes can lead to discriminative decision procedures. (2013). In: B. Custers, T. Calders, B. Schermer,

T. Zarsky (eds.) Discrimination and Privacy in the Information Society, pp. 43:57. — Caruana, R., Kangarloo, H., David, J., Dionisio, N., Sinha, U., Johnson, D. (1999). Case-based explanation of non-case-based learning methods. In: Proceedings of the 1999 American Medical Informatics Association (AMIA) Symposium, pp. 212:215.

— Chouldechova, S.: Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. arXiv preprint arXiv:1610.07524 (2016)

— Christin, A., Rosenblatt, A., boyd, d.: Courts and predictive algorithms. Data & Civil

— Citron, D., Pasquale, F.: The scored society. Washington Law Review 89(1), 1{33

— Corbett-Davies, S., Pierson, E., Feller, A., Goel, S., Huq, A.: Fair algorithms and the equal treatment principle. Working Paper (2017)

— Crawford, K., Schultz, J.: Big data and due process: Toward a framework to redress predictive privacy harms. Boston College Law Review 55(1), 93{128 (2014)

— Datta, A., Tschantz, M.C.: Automated experiments on ad privacy settings. In: Proceedings on Privacy Enhancing Technologies, pp. 92{112 (2015)

Rights Primer (2015)

(2014)

— Diakopoulos, N.: Algorithmic

pp. 214-226. ACM (2012)

University Press (2000)

accountability: Journalistic investigation of computational power structures. Digital Journalism (2015)

— Dwork, C., Hardt, M., Pitassi, T., Reingold, O., Zemel, R.: Fairness throug awareness. In: Proceedings of the 3rd Innovations in Theoretical Computer Science Conference,

— Dworkin, R.: Sovereign virtue: The theory and the practice of equality. Harvard

— Feldman, M., Friedler, S., Moeller, J., Scheidegger, C., Venkatasubramanian, S.: Certifying and removing disparate impact. In: Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 259{268 (2015)

— Fiske, Susan. Stereotyping, prejudice, and discrimination. The handbook of social psychology. (1998)

— Friedler, S.A., Scheidegger, C., Venkatasubramanian, S.: On the (im)possibility of fairness. arXiv preprint arXiv:1609.07236 (2016)

— Gillespie, T.: The relevance of algorithms. In: T. Gillespie, P. Boczkowski, K. Foot (eds.) Media technologies: Essays on communication, materiality, and society, pp. 167-193. MIT Press (2014)

— Hajian, S., Bonchi, F., Castillo, C.: Algorithmic bias: From discrimination discovery to fairness-aware data mining. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 2125{2126. ACM (2016)

— Hardt, M., Megiddo, N., Papadimitriou, C., Wootters, M.: Strategic classification. In: Proceedings of the 2016 ACM Conference on Innovations in Theoretical Computer Science, pp. 111{122. ACM (2016)

— Hardt, M., Price, E., Srebro, N.: Equality of opportunity in supervised learning. In: Proceedings of the International on Advances in Neural Information Processing Systems (NIPS), pp. 3315{3323 (2016)

— Joseph, M., Kearns, M., Morgenstern, J., Neel, S., Roth, A.: Rawlsian fairness for machine learning. arXiv preprint arXiv:1610.09559 (2016)

— Kamiran, F., Calders, T., Pechenizkiy, M.: Discrimination aware decision tree learning. In: Proceedings of 2010 IEEE International Conference on Data Mining, pp. 869-874. IEEE (2010)

— Kamishima, T., Akaho, S., Asoh, H., Sakuma, J.: Fairness-aware classifier with prejudice remover regularizer. In: Proceedings of the European Conference on Machine Learning and Principles of Knowledge Discovery in Databases (ECMLPKDD), Part II, pp. 35-50 (2011)

— Khandani, A.E., Kim, A.J., Lo, A.W.: Consumer credit risk models via machine-learning algorithms. Journal of Banking and Finance 34, 2767{2787 (2010)

— Khanna, P.: Technocracy in America: Rise of the info-state. CreateSpace Independent Publishing Platform (2017)

— Kleinberg, J., Mullainathan, S., Raghavan, M.: Inherent trade-offs in the fair determination of risk scores. In: Proceedings of the 8th Innovations in Theoretical Computer Science Conference. ACM (2017)

— Kroll, J.A., Huey, J., Barocas, S., Felten, E.W., Reidenberg, J.R., Robinson, D.G., Yu,

H.: Accountable algorithms. University of Pennsylvania Law Review 165 (2017) — Lepri, B., Staiano, J., Sangokoya, D., Letouze, E., Oliver, N.: The tyranny of data? The bright and dark sides of data-driven decision-making for social good. In: Cerquitelli T., Quercia D., Pasquale F. (eds) Transparent Data Mining for Big and Small Data. Studies in Big Data, vol 32. Springer, 2017

— Lepri, B., Oliver, N., Letouzé, E. et al. “Fair, transparent and accountable algorithmic decision-making processes. The premise, proposed solutions and open challenges”, Philos. Technol. (2017). https://doi.org/10.1007/s13347-017-0279-x

— Lipton, Z.C.: The mythos of model interpretability. In: 2016 ICMLWorkshop on Human Interpretability in Machine Learning (2016)

— Macnish, K.: Unblinking eyes: The ethics of automating surveillance. Ethics and Information Technology 14(2), 151{167 (2012)

— O'Neil, C.: Weapons of math destruction: How big data increases inequality and threatens democracy. Crown (2016)

— Pager, D., Shepherd, H.: The sociology of discrimination: Racial discrimination in employment, housing, credit and consumer market. Annual Review of Sociology 34, 181-209 (2008)

— Pasquale, F.: The Black Blox Society: The secret algorithms that control money and information. Harvard University Press (2015)

— Pedreschi, D., Ruggieri, S., Turini, F.: Discrimination-aware data mining. In: Proceedings of the 14th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 560{568 (2008)

— Ramirez, E., Brill, J., Ohlhausen, M., McSweeny, T.: Big data: A tool for inclusion or exclusion? Tech. rep., Federal Trade Commission (2016)

— Ribeiro, M., Singh, S., Guestrin, C.: “why should I trust you?": Explaining the predictions of any classifier. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pp. 1135{1144 (2016)

— Samuelson, W., Zeckhauser, R.: Status quo bias in decision making. Journal of Risk and Uncertainty (1), 7{59 (1988)

— San Pedro, J., Proserpio, D., Oliver, N.: Mobiscore: Towards universal credit scoring from mobile phone data. In: Proceedings of the International Conference on User Modeling, Adaptation and Personalization (UMAP), pp. 195{207 (2015)

— Sandvig, C., Hamilton, K., Karahalios, K., Langbort, C.: Auditing algorithms: Research methods for detecting discrimination on internet platforms. In: Data and Discrimination: Converting Critical Concerns into Productive Inquiry, a preconference at the 64th Annual Meeting of the International Communication Association (2014)

— Simonyan, K., Vedaldi, A., Zisserman, A.: Deep inside convolutional networks: Visualising image classi_cation models and saliency maps. arXiv preprint arXiv:1312.6034 (2013)

— Sunstein, C.: Regulation in an uncertain world. National Academy of Sciences (2012). URL https://www.whitehouse.gov/sites/default/_les/omb/inforeg/speeches/regul ation-in-an-uncertain-world-06202012.pdf

— Tobler, C.: Limits and potential of the concept of indirect discrimination. Tech. rep., European Network of Legal Experts in Anti-Discrimination (2008)

— Tverksy, A., Kahnemann, D.: Judgment under uncertainty: Heuristics and biases. Science 185(4157), 1124{1131 (1974)

— Wang, T., Rudin, C., Wagner, D., Sevieri, R.: Learning to detect patterns of crime. In: Machine Learning and Knowledge Discovery in Databases, pp. 515{530. Springer (2013)

— Zafar, M.B., Martinez, I.V., Rodriguez, M.D., Gummadi, K.P.: Learning fair classifiers. arXiv preprint arXiv:1507.05259 (2015)

— Zarsky, T.: The trouble with algorithmic decisions: An analytic road map to examine efficiency and fairness in automated and opaque decision making. Science, Technology and Human Values 41(1), 118{132 (2016)

— Zemel, R., Wu, Y., Swersky, K., Pitassi, T., Dwork, C.: Learning fair representation. In: Proceedings of the 2013 International Conference on Machine Learning (ICML), pp. 325-333 (2012**)**

23 Quantum Computing and Machine Learning

Antonio Puertas Gallardo

Joint Research Centre, European Commission

Wide sectors of the industry and world economy are demanding more computing power and those needs are actually of a new king of computation. A growing request for High Performance Computing (or supercomputing) power exist amidst many areas (Finance, Chemistry, Pharma-industry, Nuclear fusion research), Big data and Artificial Intelligence, in general. The more digitalization of the economy increases, the higher is the request for a bigger and different type of supercomputing. Increasingly more systems and devices are clustering together, collecting data, in what is called the dawn of the Internet of things (IoT). Artificial intelligence processors are discovering mind-blowing levels of correlations or formulating inferences in huge amount of data, but still there are plenty of signals that a considerable numbers of companies are looking for new supercomputing paradigms. Classical computers (and supercomputer) are very big calculators that performed very well doing calculus and analytics using step-by-step operations, however quantum computing will be focused on the solution of problems from a more complex and higher point of view.

Figure 7. Areas of interest for Quantum Computing¹⁸

The capacity of data stored worldwide is increasing by 20 % on a yearly basis (nowadays is ranging in the order of hundredths of Exabytes) and is a compelling force to discover new approaches to Artificial Intelligence (Machine Learning). An encouraging new concept in computation is been now investigated by the most prominent IT companies research laboratories and Academic world, is the forthcoming and hypothetical utilization of quantum computing for the optimization of the algorithms of classic machine learning. Quantum computing will not render the classic computers inappropriate. Personal computers, notebooks and smartphones will still be running on silicon-processors for the likely future and the changeover may possibly take several years. Quantum computing might be the boosting element of the new "Fourth Industrial Revolution", likewise it might be, for example, a driver for the development of new molecules for drugs, the discovering of new materials and boost Machine learning algorithms that could not have been developed before with traditional computers.

Source IBM

23.1 What is quantum computing?

A classical computer encodes information in the elementary unit of a logical bit, which can take values either "0" or "1", this information is stored and processed in the way of strings of bits (binary bits). Those individuals bit can have only one of two values: either 0 or 1. A quantum computer encodes information in the so called quantum bits or "qubits" each of them can simultaneously encode both logical bits "0" and "1" at once. This behaviour makes a quantum computer intrinsically parallel. The way to storage and process information in parallel, make some mathematical operations exponentially wide faster related to the computational speeds of classical computers for solving the same kind of problems. Quantum computation exploits a quantum physics phenomenon called "superposition" which allows to a qubit (or quantum system in general) to be in a superposition of more than one state (not only "0" or "1" as conventional computers) at the same time. The differences between classical and quantum computers can be explained with the help of a coin. In classical computing, information is stored in bits with two states, either 0 or 1 – (or heads or tails). In quantum computing, information is stored in quantum bits ("qubits") that can be any state between 0 and 1 – similar to a spinning coin that can be both heads and tails at the same time. Among other advantages, a quantum computer makes computations by the manipulation of subatomic particles. These operations are faster and with lower energy consumption if compared with the classical computers. Nowadays, the methods and instruments of quantum algorithms are very well founded and encompass a high amount of remarkable models and standards that overcome and beat the best established classical methods (See figure 4). The achievement of quantum computing is arising with IBM and Righetti succeeding in making their quantum computers available on the cloud. Many people are convinced that is only a matter of time until several theoretical designs can be tested on real-life machines. The innovative research discipline of Quantum machine Learning might offer the possibility to disrupt future approaches of intelligent data processing.

Figure 8. Computing science domains¹⁹

23.2 What is the holy grail of quantum computing?

Exponential acceleration

In other words, a quantum computer would be able to compute at a much faster speed (exponentially faster) than a classical computer. This implies that classical algorithms,

Quantum computing – weird science or the next computing revolution Morgan Stanley Research Report, August 2017

which would take years to solve on a current supercomputer, could take just hours or minutes on a quantum computer.

23.3 Quantum Machine Learning

Quantum computation and quantum information have enabled us to think physically about computation, and this approach has yielded many new and exciting capabilities for information processing. Hence, it is possible to enable us to think physically (from a quantum physics point of view) about machine learning, especially about neural networks. The field of quantum machine learning explores how to devise and implement quantum software that could allow machine learning to perform faster than on classical computers. Quantum machine learning "QML" is the science and technology at the intersection of quantum information processing and machine learning. To figure out the scientific research and work on quantum machine learning we need to consider it as a highway. On one side (one way), machine learning assists physicists to control and manipulate quantum effects and phenomena in labs. On the other side, quantum physics improves the implementation and performance of machine learning. In quantum machine learning, quantum algorithms are developed to solve typical problems of machine learning using the efficiency of quantum computing. This is usually done by adapting classical algorithms or their expensive subroutines to run on a potential quantum computer. The expectation is that in the near future, such machines will be commonly available for applications and can help to process the growing amounts of global information. The emerging field also includes approaches vice versa, namely well-established methods of machine learning that can help to extend and improve quantum information theory.

Quantum learning algorithms have been realized in a host of experimental systems and cover a range of applications as:

• Simultaneous spoken digit and speaker recognition and chaotic time-series prediction at data rates beyond a gigabyte per second (Brunner at al., 2013).
• Neural networks have been realized using liquid state nuclear magnetic resonance (Neigovzen et al., 2009).
• Defaulting on a chain of trapped ions, simulated a neural network with induced long range interactions (Pons et al., 2007).
• Solving a Higgs optimization problem with quantum annealing for machine learning (Mott et al., 2017).
• Quantum annealing versus classical machine learning applied to a simplified computational biology problem (Richard Y. Li, Rosa Di Felice et al, 2018)

23.4 Challenges

For the time being Quantum computing is still in transition between the Labs and the testing phase. This period is for the world of scientists and industry to focus on getting quantum-ready and to create a quantum-literate community who speaks quantum information language²⁰

Artificial neural networks and machine learning have now reached a new era after several decades of improvement where applications are to explode in many fields of science, industry, and technology. The Emergent Quantum information technologies would eventually boost the impact on Artificial Intelligence.

1. • Machine learning algorithms training times could be accelerated exponentially²¹.
2. • Parallelization of codes would be the new normal. ²⁰https://www.symmetrymagazine.org/article/learning-to-speak-quantum https://www.youtube.com/watch?v=Q4xBlSi_fOs
3. • Software development would be revolutionized as programmers should need to learn to make codes which manage all solutions at the same time (instantaneously, when the algorithms are deployed into the Hardware layer).

References

— Jacob Biamonte, Peter Wittek, Nicola Pancotti, Patrick Rebentrost, Nathan Wiebe, Seth Lloyd. (2017). Quantum Machine Learning. Nature 549, 195-202. https://arxiv.org/pdf/1611.09347v1.pdf

— Daniel Brunner, Miguel C. Soriano, Claudio R. Mirasso, and Ingo Fischer. Parallel photonic information processing at gigabyte per second data rates using transient states. Nat. Commun., 4:1364, January 2013.

— Dong-Ling Deng, Xiaopeng Li and S. Das Sarma. (2017). Machine learning topological states. Phys. Rev. B 96, 195145.

— Lov K. Grover. (1996). A fast quantum mechanical algorithm for database search. Quantum Physics. https://arxiv.org/abs/quant-ph/9605043

— K. Mills, M. Spanner, I. Tamblyn. Deep learning and the Schrödinger equation. Phys. Rev. A 96, 042113. https://arxiv.org/abs/1702.01361

— Marisa Pons, Veronica Ahufinger, Christof Wunderlich, Anna Sanpera, Sibylle Braungardt, Aditi Sen(De), Ujjwal Sen, and Maciej Lewenstein. Trapped ion chain as a neural network: Error resistant quantum computation. Phys. Rev. Lett., 98:023003, January 2007.

— Peter W. Shor. (1997). Polynomial-Time Algorithms for Prime Factorization and Discrete Logarithms on a Quantum Computer SIAM J. Comput., 26(5), 1484–1509. http://epubs.siam.org/doi/10.1137/S0097539795293172

24 Conclusions and future work

Emilia Gómez, Vicky Charisi, Bertin Martens, Marius Miron, Songül Tolan

Joint Research Centre, European Commission

This report has summarized the content of the 1^st workshop on Human Behaviour and Machine Intelligence (HUMAINT), which provides an interdisciplinary view on the main challenges related to the study of the impact that machine intelligence will have on human behaviour and potential needs for policy intervention.

During the workshop, we have identified several research challenges and directions that can be summarized in the following ten points:

1. There are many fundamental differences between human and machine intelligence: consciousness, evolutionary history, embodiment, situated cognition and social intelligence. In fact, we often lack of a critical approach of intelligence: what it is, its role in characterising humanness, the articulation between intelligence, knowledge and power. For instance, intelligence is much more easily granted to machines than to humans in the current media landscape. Although there are major scientific advancements on the human brain and its computational modelling, this is an extremely complex endeavour and still far from current computational models used in AI application.
2. There is not yet a full understanding of the inner workings of state-of-the-art deep neural networks. As a consequence, estimation errors might be unintuitive for humans and generalization capabilities cannot be assessed. This limits the scientific understanding of algorithms, the capability to recover from adversarial examples, and complicates human supervision in practical applications. It also raises serious questions regarding whether the goal of building humanlike intelligence is possible and desirable. We should monitor AI advancements and new computing paradigms (e.g. quantum computing).
3. We need ways to evaluate what AI can do today and predict its potential future capabilities. We need to define evaluation frameworks that are meaningful and in naturalistic settings to match practical application contexts. In this respect, we should consider engineering best practices, impact assessment methods, user satisfaction and business metrics, in order to develop smart and transparent benchmarking strategies. We should train the next generation of machine learning developers to apply and communicate these strategies and follow best practices to AI evaluation.
4. We need to advance on the explainability, accountability and transparency of algorithms in general and deep learning architectures in particular, both from a machine learning research perspective (including theoretical understanding and empirical evaluation) and from a user perspective, when these methods are exploited in a particular application context. Humans should develop a critical thinking with respect to machine intelligence, and in order to do that people need to achieve data and algorithm literacy, so that everyone can understand and challenge it.
5. With respect to human vs machine intelligence, we should move from a competition to an interaction paradigm where we should research on best strategies for collaboration and synergies exploitation between both intelligences. For instance, we need to investigate on how biases can be identified in human behaviour and if algorithms could help to recover and correct this. Here we need to consider human(owner/developer)-machine-human(user) interactions, since machine intelligence is in fact a product of human intelligence.
6. It is important to understand the interaction in the context of decision making, e.g. considering bias present in algorithms and humans and how machines can be used to overcome human bias rather than incorporate it. This is particularly relevant in

domains where decision making affects human welfare, e.g. in recruitment processes or the allocation of public funds. Also, we need to address how machines can affect human attention and strategies for humans to trust machines.