La capacité à reconnaître les visages de nos collègues, de nos amis et de nos proches est essentielle à notre réussite en tant qu'êtres sociaux. Notre cerveau accomplit cet exploit facilement et rapidement, dans une série d’opérations se déroulant en quelques dizaines de millisecondes à travers un vaste réseau cérébral du système visuel ventral. L’habileté à reconnaître les visages, par contre, varie considérablement d’une personne à l’autre. Certains individus, appelés «super-recognisers», sont capables de reconnaître des visages vus une seule fois dans la rue des années plus tôt. D’autres, appelés «prosopagnosiques», sont incapables de reconnaître le visage de leurs collègues ou leurs proches, même avec une vision parfaite. Une question simple reste encore largement sans réponse : quels mécanismes expliquent que certains individus sont meilleurs à reconnaître des visages? Cette thèse rapporte cinq articles étudiant les mécanismes perceptifs (articles 1, 2, 3) et cérébraux (articles 4, 5) derrière ces variations à travers différentes populations d’individus. L’article 1 décrit le contenu des représentations visuelles faciales chez une population avec un diagnostic de schizophrénie et d’anxiété sociale à l’aide d’une technique psychophysique Bubbles. Nous révélons pour la première fois les mécanismes en reconnaissance des expressions de cette population: un déficit de reconnaissance est accompagné par i) une sous-utilisation de la région des yeux des visages expressifs et ii) une sous-utilisation des détails fins. L’article 2 valide ensuite une nouvelle technique permettant de révéler simultanément le contenu visuel dans trois dimensions psychophysiques centrales pour le système visuel — la position, les fréquences spatiales, et l’orientation. L’article 3 a mesuré, à l'aide de cette nouvelle technique, le contenu représentationnel de 120 individus pendant la discrimination faciale du sexe et des expressions ( >500,000 observations). Nous avons observé de fortes corrélations entre l’habileté à discriminer le sexe et les expressions des visages, ainsi qu'entre l’habileté à discriminer le sexe et l’identité. Crucialement, plus un individu est habile en reconnaissance faciale, plus il utilise un contenu représentationnel similaire entre les tâches. L’article 4 a examiné les computations cérébrales de super-recognisers en utilisant l’électroencéphalographie haute-densité (EEG) et l’apprentissage automatique. Ces outils ont permis de décoder, pour la première fois, l’habileté en reconnaissance faciale à partir du cerveau avec jusqu’à 80% d’exactitude –– et ce à partir d’une seule seconde d’activité cérébrale. Nous avons ensuite utilisé la Representational Similarity Analysis (RSA) pour comparer les représentations cérébrales de nos participants à celles de modèles d’apprentissage profond visuels et langagiers. Les super-recognisers, comparé aux individus avec une habileté typique, ont des représentations cérébrales plus similaires aux computations visuelles et sémantiques de ces modèles optimaux. L’article 5 rapporte une investigation des computations cérébrales chez le cas le plus spécifique et documenté de prosopagnosie acquise, la patiente PS. Les mêmes outils computationnels et d’imagerie que ceux de l’article 4 ont permis i) de décoder les déficits d’identification faciale de PS à partir de son activité cérébrale EEG, et ii) de montrer pour la première fois que la prosopagnosie est associée à un déficit des computations visuelles de haut niveau et des computations cérébrales sémantiques. / The ability to recognise the faces of our colleagues, friends, and family members is critical to our success as social beings. Our brains accomplish this feat with astonishing ease and speed, in a series of operations taking place in tens of milliseconds across a vast brain network of the visual system. The ability to recognise faces, however, varies considerably from one person to another. Some individuals, called "super-recognisers", are able to recognise faces seen only once years earlier. Others, called "prosopagnosics", are unable to recognise the faces of their colleagues or relatives, even with perfect vision and typical intelligence. A simple question remains largely unanswered: what mechanisms explain why some individuals are better at recognizing faces? This thesis reports five articles studying the perceptual (article 1, 2, 3) and neural (article 4, 5) mechanisms behind these variations across different populations of individuals. Article 1 describes the content of visual representations of faces in a population with a comorbid diagnosis of schizophrenia and social anxiety disorder using an established psychophysical technique, Bubbles. We reveal for the first time the perceptual mechanisms of expression recognition in this population: a recognition deficit is accompanied by i) an underutilization of the eye region of expressive faces and ii) an underutilization of fine details. Article 2 then validates a new psychophysical technique that simultaneously reveals the visual content in three dimensions central to the visual system — position, spatial frequencies, and orientation. We do not know, however, whether skilled individuals perform well across a variety of facial recognition tasks and, if so, how they accomplish this feat. Article 3 measured, using the technique validated in article 2, the perceptual representations of 120 individuals during facial discrimination of gender and expressions (total of >500,000 trials). We observed strong correlations between the ability to discriminate gender and facial expressions, as well as between the ability to discriminate gender and identify faces. More importantly, we found a positive correlation between individual ability and the similarity of perceptual representations used across these tasks. Article 4 examined differences in brain dynamics between super-recognizers and typical individuals using high-density electroencephalography (EEG) and machine learning. These tools allowed us to decode, for the first time, facial recognition ability from the brain with up to 80% accuracy — using a mere second of brain activity. We then used Representational Similarity Analysis (RSA) to compare our participants' brain representations to those of deep learning models of object and language classification. This showed that super-recognisers, compared to individuals with typical perceptual abilites, had brain representations more similar to the visual and semantic computations of these optimal models. Article 5 reports an investigation of brain computations in the most specific and documented case of acquired prosopagnosia, patient PS. The same computational tools used in article 4 enabled us to decode PS's facial identification deficits from her brain dynamics. Crucially, associations between brain deep learning models showed for the first time that prosopagnosia is associated with deficits in high-level visual and semantic brain computations.

No description available.

La reconstruction faciale permet d’approximer un visage sur la base d’un crâne, lorsque des restes humains sont retrouvés. En science forensique, elle est l’un des outils utilisés dans un but d’identification post-mortem. Les procédures actuelles d’approximation ne sont pas standardisées et constamment revisitées. Il est également possible d’obtenir des prédictions du phénotype d’un individu (caractères physiques apparents) à partir de son ADN, qui pourraient être ajoutées aux reconstructions pour maximiser les chances de reconnaissance. Cette recherche vise à étudier l’approximation faciale à des fins de reconnaissance et l’apport du phénotypage par l’ADN à cette pratique. Le crâne et le relevé biologique d’un donneur du laboratoire d’anatomie de l’UQTR ont été utilisés. Six praticiens ont approximé son visage à partir d’une copie de son crâne et de ses données phénotypiques et anthropologiques. Les résultats corroborent qu’il existe un manque de standardisation des méthodes et techniques, menant à différents résultats selon le praticien. Des tests de reconnaissance et de ressemblance ont été effectués à l’aide d’un échantillon de 46 participants. Malgré la variabilité des approximations, elles ont toutes été reconnues au moins une fois lors des tests de reconnaissance, soulevant la possibilité que la reconnaissance d’un visage est idiosyncratique. Les caractéristiques qui semblent influencer davantage la reconnaissance sont la forme, la taille et la position des yeux, de la bouche et du nez. Finalement, au regard des incertitudes propres à la méthode et du rôle investigatif de l’approximation faciale, il est recommandé que le phénotypage accompagne l’accompagne par écrit. / Facial reconstruction is a process by which a face is approximated from a skull when human remains are found. In forensic science, it is one of the many tools used for the purpose of post-mortem identification. The current approximation procedures are not standardized and are always revisited. Nowadays, it is possible to obtain phenotype (apparent physical traits) predictions from an individual’s DNA. These predictions could be added to facial approximations to maximize the chances of recognition. This research aims to study facial approximation for recognition purposes and the plus-value of DNA phenotyping to facial approximation. The skull and biological material from one donor of the UQTR’s anatomy laboratory were used. Six practitioners approximated the donor’s face using a copy of his skull, and phenotyping and anthropological data. The results corroborated the lack of standardization regarding the approximating methods and techniques, which leads to different resulting approximations depending on the practitioner. Recognition and resemblance tests were carried out with a sample of 46 participants. Despite the wide variability of the approximations, they were all recognized at least once during the recognition tests, raising the possibility that the recognition of a face is idiosyncratic. The characteristics that seemed the most important to recognition were the shape, size and position of the eyes, the mouth, and the nose. Finally, with regard to the uncertainties specific to the method and the final investigative role of facial approximation, phenotyping would benefit in feeding a spoken portrait.

Facial recognition technology has become a ubiquitous tool in security and personal identification. However, the rise of this technology has been accompanied by concerns over inherent biases, particularly regarding ethnic and gender. This thesis examines the extent of these biases by focusing on the influence of dataset imbalances in facial recognition algorithms. We employ a structured methodological approach that integrates AI-generated images to enhance dataset diversity, with the intent to balance representation across ethnics and genders. Using the ResNet and Vgg model, we conducted a series of controlled experiments that compare the performance impacts of balanced versus imbalanced datasets. Our analysis includes the use of confusion matrices and accuracy, precision, recall and F1-score metrics to critically assess the model’s performance. The results demonstrate how tailored augmentation of training datasets can mitigate bias, leading to more equitable outcomes in facial recognition technology. We present our findings with the aim of contributing to the ongoing dialogue regarding AI fairness and propose a framework for future research in the field.

The Wechsler Memory Scale –iii is the newest version of a six-decade old neuropsychological inventory. Since its conception, the Wechsler Memory Scale has been highly utilized by practitioners to accurately assess various memory functions in adult subjects. Revisions made within this inventory include the Faces I subtest, a facial recognition scale, which was added in order to strengthen the instrument’s accuracy at measuring episodic memory. Facial recognition, both cross-race and within-race, has been researched extensively and consistent biases have been found between race of test taker and cross-racial identification. Theories of exposure/contextual interaction (environment) and biological foundations have been the subject of study in the past in order to determine from where these racial identification deficits stem. The current study focuses on revealing bias in the Faces I subtest, regarding to an unequal distribution of racially representative faces in the testing materials. Eighty-eight college students were recruited to view forty-eight pictured faces from the Faces I subtest and determine the racial category to which the pictured face belonged. The subjects’ categorical responses were the basis for calculating a percent agreement score for racial category of each face. It was determined, using the results of subjects’ responses, that the Faces I subtest contained an unequal distribution of racially representative faces in both the Target and Interference testing material. This confirmed the presence of an inherent bias within the subscale. The implications of memory accuracy for the WMS-iii are discussed as it relates to different fields of study, but none more directly than the criminal justice system. Eyewitness testimony is a pivotal evidentiary tool in the criminal justice system, and ramifications of cross-racial identification deficits and biases in the tools to accurately assess memory are increasingly bringing this once heavily relied upon tool into question.

The topic of this thesis is AI recruitment chatbots, digital discrimination, and data feminism (D´Ignazio and F.Klein 2020), where I aim to critically analyze issues of bias in these types of human-machine interaction technologies. Coming from a professional background of theatre, performance art, and drama, I am curious to analyze how using AI and social robots as hiring tools entails a new type of “stage” (actor’s space), with a special emphasis on social acting. Humans are now required to adjust their performance and facial expressions in the search for, and approval of, a new job. I will use my “theatrical glasses” with an intersectional lens, and through a methodology of cultural analysis, reflect on various examples of conversational AI used in recruitment processes. The silver bullet syndrome is a term that points to a tendency to believe in a miraculous new technological tool that will “magically” solve human-related problems in a company or an organization. The captivating marketing message of the Swedish recruitment conversational AI tool – Tengai Unbiased – is the promise of a scientifically proven objective hiring tool, to solve the diversity problem for company management. But is it really free from bias? According to Karen Barad, agency is not an attribute, but the ongoing reconfiguration of the world influenced by what she terms intra-actions, a mutual constitution of entanglement between human and non-human agencies (2003:818). However, tech developers often disregard their entanglement of human-to-machine interferences which unfortunately generates unconscious bias. The thesis raises ethical questions of how algorithmic measurement of social competence risks holding unconscious biases, benefiting those already privileged or those acting within a normative spectrum.

No description available.