Projet de thèse en Sciences cognitives
Sous la direction de Nicolas Baumard.
Thèses en préparation à Paris Sciences et Lettres , dans le cadre de École doctorale Cerveau, cognition, comportement (Paris) , en partenariat avec Institut Jean-Nicod (Paris) (laboratoire) et de Ecole normale supérieure (établissement de préparation de la thèse) depuis le 01-09-2018 .
Un certain nombre d'observations historiques suggèrent que la confiance sociale a clairement augmenté en Europe, au moins depuis la fin du Moyen-Âge : les guerres de religions et chasses aux sorcières ont disparu, les duels et vengeances ont perdu leur attrait, la liberté intellectuelle est devenue une valeur centrale des pays modernes, etc. (McCloskey, 2016). Depuis le travail avant-gardiste de Norbert Elias, "The Civilizing Process", les historiens ont utilisés les manuels de l'étiquette, les registres des clubs les normes de réciprocité et d'échanges de cadeaux pour documenter cette augmentation de la confiance (Clark, 2000; Sunderland, 2007). Cependant, ces observations sont qualitatives et limitées à des domaines spécifiques et des périodes historiques assez courtes. On ne peut donc pas les utiliser pour comparer la confiance sociale entre les sociétés et au cours du temps et pour distinguer les différentes causes de l'augmentation de la confiance. Ce projet a pour objectif d'appliquer les outils et connaissances de la cognition sociale pour quantifier et expliquer l'augmentation de la confiance dans les sociétés passées.
Signals of Trust in History: Can Social Cognition Inform Cultural Evolution?
A number of historical observations suggest that social trust steadily increased in Europe from at least the late Middle Ages onwards: wars of religions and witch hunts abated, honour killings and revenge lost their appeal, intellectual freedom became a central value of modern countries, etc. (McCloskey, 2016). Since the pioneering work of Norbert Elias, The Civilizing Process, historians have used etiquette manuals, registries of friendly societies and clubs, norms of reciprocity and gift exchanges to document the rise of social trust (Clark, 2000; Sunderland, 2007). However, such observations are qualitative and limited to very specific domains and short historical periods. As a result, they cannot be used to compare social trust between societies and across times, and to disentangle the various causes of the rise of trust. The aim of the proposed project is thus to apply the tools and insights of social cognition to quantify and explain the rise of trust in past societies. Quantifying trust in past societies In psychology and economics, social trust is usually measured by asking participants questions such as: Generally speaking, would you say that most people can be trusted or that you need to be very careful in dealing with people?' (World Value Survey). Recently, behavioural scientists have also relied on economic games such as the Trust Game or the Public Goods Game, using real incentives to test whether people are willing to trust others in a situation of cooperation. Quite obviously, we cannot go back in time and ask people to fill questionnaires or play economic games but we still have access to what their minds produced: books, songs, paintings, sculptures, etc. These cultural artefacts are the remnants of people's past psychologies and can function as cognitive fossils of extinct mentalities and preferences. Recent work has indeed shown that people's preferences in various areas of social cognition are reflected in cultural artefacts: Costa and Corazza (2006) demonstrated that the people's preference for friendly-looking faces leads painters to exaggerate neotenic features in their portraits (big eyes or round faces). Similarly, fictions, such as romance novels (Salmon, 2012), TV shows (Fisher, 2012), epic poems (Gottschall, 2008) or tragedies (Nettle, 2005), are all consistently aligned with humans' universal interest for information related to mating, commitment and status competition (for reviews and discussions, see Gottschall & Wilson, 2005; Pinker, 2007). Moving away from cognitive sciences, historians have long relied on narratives and portraits to track changes in mentalities (see for instance McCloskey (2016)). This body of research, however, is often impressionistic. Figure 1 | Trustworthiness and dominance inferences from facial cues. Findings in experimental psychology indicate that specific trait inferences can be represented within a 2D space defined by subjective evaluations of valence/trustworthiness and power/dominance in faces. (a) Features that appear untrustworthy include low inner eyebrows, a deep nasal indentation, shallow cheekbones and a thin chin. (b) Features that appear trustworthy include high inner eyebrows, a shallow nasal indentation, pronounced cheekbones and a wide chin. The aim of this project is therefore to apply recent tools developed in social cognition to provide more accurate estimates of the rise and fall of social trust across time. Specifically, experimental work in social cognition has demonstrated that human faces display consistent cues of trustworthiness and dominance (Todorov, Said, Engell, & Oosterhof, 2008) (Figure 1). In the same way, emotions can be measured in written texts. For instance, Using WordNet, a standard tool in text mining (Strapparava & Mihalcea, 2008), Acerbi et al. (2013) leveraged the dataset provided by Google Ngram to study cultural trends in the use of mood words and found evidence for distinct historical periods of positive and negative moods. Study 1. In The Smile Revolution, cultural historian Colin Jones notes that a rise of smiling portraits took place in 18th century France, with a concomitant shift in politeness, sensibility and self-presentation (Jones, 2014). WP1 aims to systematically document these qualitative observations and to quantify the increase of trust in portraits. In order to estimate people's preferences, we will use two methods to measure perceived trustworthiness and perceived dominance in portraits: a subjective estimation based on participants' ratings and an automated estimation based on the detection of facial action units. Subjective estimation of facial trust: Participants will be recruited online through Oxford University's testing platform, Prolific Academic. Participants will be asked to assess levels of trustworthiness and dominance in portraits. In order to remove any influence of the historical context, of painting style or fashion-related cues (such as beards, hair styles, hats, and so on), portraits will first be processed using Facegen (Figure 2), a software that has already been proven to create realistic avatars (Singilar Inversions, 2005; Verosky & Todorov, 2010). Figure 2 | Quantifying trustworthiness and dominance in historical portraits. (a) Historical portraits, here Henry VIII produced by the workshop of Hans Holbein the Younger in 1537, will be collected to measure facial cues of trustworthiness and dominance following two estimation methods. (b) The subjective estimation method will rely on online participants who will be asked to judge trustworthiness and dominance in portraits stripped of all fashion or stylistic related cues using the open software FaceGen© (Singilar Inversions, 2017). (c) The automated estimation of facial trust will be performed by training an algorithm based on overfitting linear models to predict trutstworthiness and dominance based on facial action units extracted by the open software OpenFace (Amos, Ludwiczuk, & Satyanarayanan, 2016). While this subjective method has the advantage of relying on real human judgments, it is intrinsically time-consuming: Large numbers of participants need to be tested to obtain reliable estimates and Facegen avatars are created manually, which requires the specification of a number of action units (eyes, mouth, etc.) in the algorithm for each portrait. For this reason, the subjective method cannot be used for very large databases. Automated estimation of facial trust: In order to produce human-like trustworthiness and dominance estimations, we will design an algorithm that will automatically generate trustworthiness and dominance evaluations on the basis of every facial action unit extracted by OpenFace, i.e. smile, eye brows, etc. (Baltruaitis, Robinson, & Morency, 2016). The extracted facial features will then be used to predict perceived trustworthiness and perceived dominance in faces. To do so, we will extend Oosterhof and Todorov's (2008) methodology and run overfitting linear models based on all the extracted facial action units to predict trustworthiness and dominance in validated avatars. We will use two databases of validated avatars to train the algorithm: those created by Todorov's team and the Karolinska face database (Lundqvist, Flykt, & Öhman, 1998). The model will then be validated by using its parameters to predict perceived trustworthiness and dominance of an independent set of ratings made on photographs. Once validated, the model's parameters will be used to estimate perceived trustworthiness and dominance, based on the facial features extracted by OpenFace on large historical databases of paintings. As a first step, we will use large databases in which portraits are well digitized and well indexed as portraits: Web Gallery of Arts, Joconde Catalogue, The National Portrait Gallery, The Rijksmuseum Amsterdam, The Metropolitan Museum of Art, The Prado Museum. One limitation of this first step however, is that a change in trust displays could be due to a change in the social identity of the sitters or in the nature of the portraits (e.g., private/public). In order to assess the robustness of our findings, we will therefore control for this potential bias in two follow-up analyses. First, we will focus on portraits of individuals who have all held the same social position, such as Head of State. In England, for instance, we will focus on portraits of Speakers of the House of the Commons, Chancellors of Oxford and Cambridge, Archbishops of Canterbury, Poet Laureates, and Lucasian Professors of Mathematics. A preliminary inquiry suggests that it is possible to go back to the beginning of the 16th century for these English series. Second, we will analyse portraits of recurrent mythological characters such as Paris in The Judgments of Paris using structured databases, e.g., Wikimedia Commons and the Aby Warburg Iconographic Database. A preliminary inquiry suggests that the most popular topics have been painted recurrently for at least 3 centuries (e.g. Figure 3). Study 2 Trust in Narrative Fictions Historians of literature have long noticed that narrative fictions reflect the preferences of their original audience. For instance, while the Greek archaic epics (The Iliad, The Odyssey) emphasized social status, physical deeds and bravery (Gottschall, 2008), the Greek novels of the Imperial period (The Ephesian Tale, Daphnis and Chloe), written in the same language, in the same place, but in a much wealthier and peaceful society, emphasized the generosity, the equanimity and the personal qualities of the heroes (Konstan, 2014). Similar observations apply to the comparison of early medieval English literature (Beowulf, Arthurian legend) and to late medieval English literature (Tristan and Iseult, Marie de France's Lays) (Reddy, 2012). Study 2 will quantify the importance of markers of trust using text-mining (Iliev, Dehghani, & Sagi, 2015) and topic modelling (Blei, 2012) in online summaries and digitized texts (Study 2a), and test whether their importance is associated with living standards. To control for possible biases in the quality of material and translations across periods, we will run additional analyses on well-controlled databases such as Théâtre Classique', in which all French plays written between 1500 and 1800 have been digitized and encoded as part of the Text Encoding Initiative (TEI). Computational literary techniques, such as topic modelling, will then be applied to digitized texts to compare texts written in poor and wealthy periods. Topic modelling algorithms are well-tried statistical methods that analyse words in order to identify the themes that are most represented in a given text (Blei, 2012). Importantly, while text mining requires that scholars know what terms or topics are worth searching, topic modelling makes no assumption. The topics emerge automatically from the analysis of the original texts, thereby limiting researchers' degrees of freedom (see for instance Estill & Meneses, 2018 on the evolution of characters in Shakespeare's plays). We predict that narratives written during poor vs. wealthy times will not use the same topics, and that topics related to trust and cooperation will be more important in periods of affluence. References Acerbi, A., Lampos, V., Garnett, P., & Bentley, R. A. (2013). The expression of emotions in 20th century books. PloS one, 8(3), e59030. Baltruaitis, T., Robinson, P., & Morency, L.-P. (2016). Openface: an open source facial behavior analysis toolkit. In Applications of Computer Vision (WACV), 2016 IEEE Winter Conference on (p. 110). IEEE. Clark, P. (2000). British clubs and societies 1580-1800: the origins of an associational world. OUP Oxford. Costa, M., & Corazza, L. (2006). Aesthetic phenomena as supernormal stimuli: The case of eye, lip, and lower-face size and roundness in artistic portraits. Perception, 35(2), 229246. Fisher, M. L. (2012). Why who shot JR Matters: Dallas as the pinnacle of human evolutionary television. Review of General Psychology, 16(2), 200. Fouquet, R., & Broadberry, S. (2015). Seven centuries of European economic growth and decline. The Journal of Economic Perspectives, 29(4), 227244. Gottschall, J. (2008). The rape of Troy: Evolution, violence, and the world of Homer. Cambridge University Press. Gottschall, J., & Wilson, D. S. (2005). The literary animal: Evolution and the nature of narrative. Northwestern University Press. Haushofer, J. (2013). The psychology of poverty: Evidence from 43 countries. Documento de Trabajo). Disponible en http://web. mit. edu/joha/www/.[Links]. Consulté à l'adresse http://www.princeton.edu/joha/publications/Haushofer_2013.pdf Hörl, M., Kesternich, I., Smith, J. P., & Winter, J. K. (2016). Early-life Circumstances Predict Measures of Trust among Adults: Evidence from Hunger Episodes in Post-War Germany. Jones, C. (2014). The smile revolution in eighteenth-century Paris. Oxford: Oxford University Press. Lundqvist, D., Flykt, A., & Öhman, A. (1998). The Karolinska directed emotional faces (KDEF). CD ROM from Department of Clinical Neuroscience, Psychology section, Karolinska Institutet, (1998). McCloskey, D. N. (2016). Bourgeois equality: How ideas, not capital or institutions, enriched the world. University of Chicago Press. Nettle, D. (2005). The wheel of fire and the mating game: Explaining the origins of tragedy and comedy. Journal of Cultural and Evolutionary Psychology, 3(1), 3956. Petersen, M. B., & Aarøe, L. (2015). Birth Weight and Social Trust in Adulthood Evidence for Early Calibration of Social Cognition. Psychological science, 26(11), 16811692. Pinker, S. (2007). Toward a consilient study of literature. Philosophy and Literature, 31(1), 162178. Salmon, C. (2012). The pop culture of sex: An evolutionary window on the worlds of pornography and romance. Review of General Psychology, 16(2), 152. Singilar Inversions. (2005). Facegen Main Software Development Kit (Singular Inversions, Vancouver, BC, Canada). Version. Strapparava, C., & Mihalcea, R. (2008). Learning to identify emotions in text. In Proceedings of the 2008 ACM symposium on Applied computing (p. 15561560). ACM. Sunderland, D. (2007). Social Capital, Trust and the Industrial Revolution: 17801880. Routledge. Todorov, A., Said, C. P., Engell, A. D., & Oosterhof, N. N. (2008). Understanding evaluation of faces on social dimensions. Trends in cognitive sciences, 12(12), 455460. Verosky, S. C., & Todorov, A. (2010). Generalization of affective learning about faces to perceptually similar faces. Psychological Science, 21(6), 779785.