HILL goes to CogSci

We are happy to announce that our publication-based talk and two posters has been accepted to 42nd Annual Virtual Meeting of the Cognitive Science Society. We will present the work concerning our research projects on linguistic development, categorization and agent-based modeling of language.

1. Structured ecologies for social and linguistic development
Joanna Rączaszek-Leonardi, Human Interactivity and Language Lab, Faculty of Psychology, University of Warsaw, Warsaw, Poland
Katharina Rohlfing, Paderborn University, Paderborn, Germany
Abstract
This is a joint work of two labs that offers a perspective on development and learning, which complements the conference’s focus on “changes in representation and processing abilities in development”. Strong background in ecological psychology allowed us to recognize the richness and multilayered structuring of infants’ environment, which actively engages them and to which infants tune their action-perception. We conceptualize this environment as reliable “social physics”, constituted of predictable, enacted social events, in which infants learn to participate. Using both traditional (qualitative and quantitative) and dynamical systems methods, we show the structuring of such events on multiple timescales and levels and how participating in them sculpts the child’s agency in the social world. We show how this background allows a fresh look on language acquisition and how it informs computational modelling of language emergence and models of human-robot interaction.

2. Abstraction and Generalization: Comparing Adaptive Models of Categorization
Julian Zubek, Human Interactivity and Language Lab, Faculty of Psychology, University of Warsaw, Warsaw, Poland
Ludmila Kuncheva, School of Computer Science and Electronic Engineering, Bangor University, Bangor, United Kingdom
Abstract
Between prototype and exemplar models of categorization lie adaptive models, which represent categories using a varying number of reference points. They regulate the amount of abstraction they make depending on the category structure. Motivated by ecological considerations, we investigate whether adopting such adaptive strategies could improve generalization in realistic environments. We compare performance of four adaptive models: RMC, SUSTAIN, REX, VAM with that of prototype and exemplar models on three artificial and three natural category structures. Both the exemplar model with adapted sensitivity parameter and VAM perform well on category structures requiring different amount of abstraction. Our results confirm the importance of the link between abstraction and generalization.

3. The Emergence of Action-grounded Compositional Communication
Krzysztof Główka, Michał Niklewski, Joanna Wiszowata, Tomasz Korbak, Joanna Rączaszek-Leonardi, Julian Zubek, Human Interactivity and Language Lab, Faculty of Psychology, University of Warsaw, Warsaw, Poland
Abstract
Classical models of the emergence of compositionality in communication focused on the compositional nature of the environment (Cangelosi, 2001; Cornish et al.,2008). Here we advance a model for compositional signal emergence from the integration of environment’s properties with agents’ actions. We take as a starting pointCangelosi’s (2001) model, where a population of agents searched for edible mushrooms. Given opportunity to communicate, they evolved a system in which combinations of signs were sensorily grounded in combinations of mushroom properties. We modify this model by grounding the communication also in agents’ actions. With this, we may evolve communication systems containing meaningful compositions of mushroom properties and agent actions. We investigate how such compositions can facilitate a) learning the communication protocol, b)learning the adequate behaviour policy. This kind of “sensory-motor” compositionality seems better suited for coordinating navigation in dynamic environments. We demonstrate that agents receiving predefined language signal systematically outperform agents with no language in noisy environments. Furthermore, we show that effective agents’ policies employ long sequential procedures, with length of the procedure translating to its effectiveness. This opens a possibility for the emergence of complex language controlling complex actions.