The emergence of a language is rarely directly observed in a natural environment. Similar to a phenomenon previously observed in Nicaragua, deaf Dominican children appear to have created a rudimentary form of manual communication in absence of comprehensible linguistic input. The evolution of this communication system over the course of five years (2007-2012) is documented as part of a cross-cultural and cross-generational study in which sign complexity is analyzed. The role of innate and environmental components of language creation and acquisition are discussed using data from hearing children and parents, including the parents of the deaf Dominican children cited above. Results confirm that a new communication system is indeed slowly emerging in the Dominican Republic, and that this system shows signs of evolution in the period extending from 2007 to 2012. Signs produced by the deaf Dominican children meet the minimal requirements for a communicative symbol, show signs of mutual intelligibility, and differ from the signs of the other implemented Sign Languages in the Dominican Republic. Two cohorts of manual communicators appear to be present, and younger signers seem to have more advanced linguistic competencies in comparison to older signers within the community. The signs that are part of the observed Dominican manual communication system also appear to differ in complexity from those produced by hearing adults and children, supporting the presence of innate abilities for language creation. Specifically, the deaf Dominican children are generally found to have more diversified sign repertoires and to display faster signing rates over time, in comparison to hearing adults and children. Qualitative data and quantitative trends further support a more complex understanding by deaf children of the use of signs as an independent communication system from speech. Analyses looking at the impact of input on language creation provides some support for the existence of infant-directed signing in a way similar to what is observed with infant-directed speech. The use of repetitions by hearing adults using infant-directed silent gestures could provide support for usage-based theories of language development. That being said, young hearing children with no prior exposure to Sign Language and with minimal relative linguistic experience were found to produce signs equivalent in complexity to those of hearing adults, therefore potentially providing further support for an innate understanding of complex linguistic rules. Deaf Dominican children were further found to surpass the input received by hearing adults over time. In all, this research is consistent with previous studies attesting for children’s natural ability for language creation and development.

No description available.

Nous apprenons très jeunes une quantité de règles nous permettant d'interagir avec d'autres personnes: des conventions sociales. Elles diffèrent des autres types d'apprentissage dans le sens où les premières personnes à les avoir utilisées n'ont fait qu'un choix arbitraire parmi plusieurs alternatives possibles: le côté de la route où conduire, la forme d'une prise électrique, ou inventer de nouveaux mots. À cause de celà, lorsqu'une nouvelle convention se crée au sein d'une population d'individus interagissant entre eux, de nombreuses alternatives peuvent apparaître et conduire à une situation complexe où plusieurs conventions équivalentes coexistent en compétition. Il peut devenir difficile de les retenir toutes, comment faisons-nous pour trouver un accord efficacement ? Nous exerçons communément un contrôle actif sur nos situations d'apprentissage, en par exemple sélectionnant des activités qui ne soient ni trop simples ni trop complexes. Il a été montré que ce type de comportement, dans des cas comme l'apprentissage sensori-moteur, aide à apprendre mieux, plus vite, et avec moins d'exemples. Est-ce que de tels mécanismes pourraient aussi influencer la négociation de conventions sociales? Le lexique est un exemple particulier de convention sociale: quels mots associer avec tel objet ou tel sens? Une classe de modèles computationels, les Language Games, montrent qu'il est possible pour une population d'individus de construire un langage commun via une série d'interactions par paires. En particulier, le modèle appelé Naming Game met l'accent sur la formation du lexique reliant mots et sens, et montre une typique explosion de la complexité avant de commencer à écarter les conventions synonymes ou homonymes et arriver à un consensus. Dans cette thèse, nous introduisons l'idée de l'apprentissage actif et du contrôle actif de la croissance de la complexité dans le Naming Game, sous la forme d'une politique de choix du sujet de conversation, applicable à chaque interaction. Différentes stratégies sont introduites, et ont des impacts différents sur à la fois le temps nécessaire pour converger vers un consensus et la quantité de mémoire nécessaire à chaque individu. Premièrement, nous limitons artificiellement la mémoire des agents pour éviter l'explosion de complexité locale. Quelques stratégies sont présentées, certaines ayant des propriétés similaires au cas standard en termes de temps de convergence. Dans un deuxième temps, nous formalisons ce que les agents doivent optimiser, en se basant sur une représentation de l'état moyen de la population. Deux stratégies inspirées de cette notion permettent de limiter les besoins en mémoire sans avoir à contraindre le système, et en prime permettent de converger plus rapidement. Nous montrons ensuite que la dynamique obtenue est proche d'un comportement théorique optimal, exprimé comme une borne inférieure au temps de convergence. Finalement, nous avons mis en place une expérience utilisateur en ligne sous forme de jeu pour collecter des données sur le comportement d'utilisateurs réels placés dans le cadre du modèle. Les résultats suggèrent qu'ils ont effectivement une politique active de choix de sujet de conversation, en comparaison avec un choix aléatoire.Les contributions de ce travail de thèse incluent aussi une classification des modèles de Naming Games existants, et un cadriciel open-source pour les simuler. / Social conventions are learned mostly at a young age, but are quite different from other domains, like for example sensorimotor skills. The first people to define conventions just picked an arbitrary alternative between several options: a side of the road to drive on, the design of an electric plug, or inventing a new word. Because of this, while setting a new convention in a population of interacting individuals, many competing options can arise, and lead to a situation of growing complexity if many parallel inventions happen. How do we deal with this issue?Humans often exhert an active control on their learning situation, by for example selecting activities that are neither too complex nor too simple. This behavior, in cases like sensorimotor learning, has been shown to help learn faster, better, and with fewer examples. Could such mechanisms also have an impact on the negotiation of social conventions ? A particular example of social convention is the lexicon: which words we associated with given meanings. Computational models of language emergence, called the Language Games, showed that it is possible for a population of agents to build a common language through only pairwise interactions. In particular, the Naming Game model focuses on the formation of the lexicon mapping words and meanings, and shows a typical burst of complexity before starting to discard options and find a final consensus. In this thesis, we introduce the idea of active learning and active control of complexity growth in the Naming Game, in the form of a topic choice policy: agents can choose the meaning they want to talk about in each interaction. Several strategies were introduced, and have a different impact on both the time needed to converge to a consensus and the amount of memory needed by individual agents. Firstly, we artificially constrain the memory of agents to avoid the local complexity burst. A few strategies are presented, some of which can have similar convergence speed as in the standard case. Secondly, we formalize what agents need to optimize, based on a representation of the average state of the population. A couple of strategies inspired by this notion help keep the memory usage low without having constraints, but also result in a faster convergence process. We then show that the obtained dynamics are close to an optimal behavior, expressed analytically as a lower bound to convergence time. Eventually, we designed an online user experiment to collect data on how humans would behave in the same model, which shows that they do have an active topic choice policy, and do not choose randomly. Contributions from this thesis also include a classification of the existing Naming Game models and an open-source framework to simulate them.

In the recent decades, alternative notions regarding the role of symbols in intelligence in natural and artificial systems have attracted significant inter- est. The main difference of the so-called situated and embodied approaches to cognitive science from the traditional cognitivist position is that symbolic repre- sentations are viewed as resources, similar to maps used for navigation or plans for activity, instead of as transparent stand-ins in internal world models. Thus, all symbolic resources have to be interpreted and re-contextualized for use in concrete situations. In this view, one of the primary sources of such symbolic resources is language. Cognitivism views language as a vessel carrying informa- tion originally located in the processing mechanisms of the individual agents. Situated approaches, on the other hand, view language both as a communicative mechanism and as a means for the individual agents to enhance and extend their cognitive machinery, by e.g. better utilizing their attentional resources, or mod- ifying their perceptual-motor means. Taking inspiration from these ideas, and building on multi-agent models developed in other fields, the field of language evolution developed models of the emergence of shared resources for communi- cation in a community of agents. In these models, agents with various means of categorization and learning engage in communicative interactions with each other, using shared signs to refer either to pre-given meanings or entities in a situation. In order to avoid falling into the same mentalist pitfalls as cognitivism in the design of these models, such as the stipulation of an inner sphere of mean- ings for which communicative signs are mere labels, the role of communication should be viewed as one of the social coordination of behavior using physically grounded symbols. To this end, an experimental setup for language games, and a robotic model for agents which engage in such games are presented. The setup allows the agents to utilize shared symbols in the completion of a simple task, with one agent instructing another on which action to undertake. The symbols used by agents in the language games are grounded in the embodied choices presented to them by their environment, and the agents can further use the symbols created in these games for enhancing their own behavioral means. The learning mechanism of the agents is similarity-based, and uses low-level sensory data to avoid the building in of features. Experiments have shown that the establishment of a common vocabulary of labels depends on how well the instructors are trained on the task and the availability of feedback mechanisms for the exchanged labels.

L'émergence du langage est considérée comme l'une des marques de l'intelligence humaine. Par conséquent, nous émettons l'hypothèse que l'émergence de latences ou de représentations similaires au langage dans un système d'apprentissage profond pourrait aider les modèles à obtenir une meilleure généralisation compositionnelle et hors distribution. Dans cette thèse, nous présentons une série d'articles qui explorent cette hypothèse dans différents domaines, notamment l'apprentissage interactif du langage, l'apprentissage par imitation et la vision par ordinateur. / The emergence of language is regarded as one of the hallmarks of human intelligence. Therefore, we hypothesize that the emergence of language-like latents or representations in a deep learning system could help models achieve better compositional and out-of-distribution generalization. In this thesis, we present a series of papers that explores this hypothesis in different fields including interactive language learning, imitation learning and computer vision.