Symbolic associations in neural network activations: representations in the emergence of communication Emerson Silva de Oliveira Cognitive and Intelligent Systems Lab (LASIC) State University of Feira de Santana (UEFS) Feira de Santana, Brazil emersonso@dcc.ufba.br Angelo Loula Cognitive and Intelligent Systems Lab (LASIC) State University of Feira de Santana (UEFS) Feira de Santana, Brazil angelocl@uefs.br Abstract— Representation has a fundamental role in Artificial Intelligence but there is still an open debate on basic issues on this subject. Particularly, there have been various studies on the emergence of communication and language in artificial agents, where the debate on representations underlying theseprocesses should be significant, however not much discussion and studies have been done. We propose to identify and classify possible representational processes occurring during the emergence of communication, replicating a computational experiment previously proposed and evaluating neural network activations patterns. To define representation and its classes, including icons, indexes and symbols, we rely on the semiotics of Charles Sanders Peirce. Results show that symbolic associations are established during the evolution of artificial agents and such symbolic associations benefit adaptive success. Keywords— symbols; representations; neural network activation; communication; semiotics. I. INTRODUCTION Representation is a topic of large debate in Artificial Intelligence since its beginning, but there is still an open debate on basic issues. Beer[1] emphasizes that, despite the fundamental role of representation in computational approaches to intelligent systems, there is no consensus on its definition. This problem can be particularly significant in discussions on language and communication in artificial agents (see, e.g., [2-5]). Various works on the emergence of communication and language in artificial agents have been developed, with technological and scientific advance motivations (see [6-7]). Nevertheless, in these works, little (or almost nothing) is discussed about possible representational processes in the emergent communication between agents. Once communication can be seen as the production of representations by an utterer and the interpretation of representation by an interpreter, we regard the discussion on representations in such scenario as an important issue and, as such, it can further advance research in this field. In the present work, we propose to identify and classify possible representational processes occurring during the emergence of communication replicating a computational experiment previously proposed. Instead of proposing a new experiment, we chose to replicate a previous one as a mean to show how computational studies on the emergence of communication can be expanded to include results and discussions on representation. And, to do so, we study the neural network activation patterns to identify symbolic associations. To define representation and its classes, comprising icons, indexes and symbols, we rely on the theory of signs of Charles Sanders Peirce. We replicate the Artificial Life experiment of Mirolli and Parisi [8], in which a community of agents controlled by neural networks is evolved to study the emergence of communication in a scenario in which communication only benefits the interpreter. In the next section, we briefly review related works, describing simulations on the emergence of communication and analysis of neural networks controlling artificial agents. In section III, we describe the semiotic theory of C.S.Peirce, which is the theoretical framework for our investigation. In section IV, we present the replicated experiment and our methodology to search and identify representations in this experiment. We finish by presenting results and conclusions, highlighting new perspectives on the study of representations in the emergence of communication. II. RELATED WORK There have been several works on computational experiments related to the emergence of communication in a community of artificial agents [6-7]. However, discussions on the underlying representation processes, particularly in those using neural networks, find little space in such literature. We will review a few representative studies that deal with the emergence of communication among agents controlled by neural networks that are relevant in the context of this work. Floreano et. al. [9] studied conditions for the emergence and development of communication in a community of robotic agents. Robots are controlled by a two-layer neural network, with inputs receiving directions of the blue and red lights around them and from a ground sensor that identifies the ground color near food and poison locations in the environment. The output is connected to the right and left motors, and also to a blue light ring on the robot. The robots are divided in groups and evolved for the task of reaching and staying in the food location, and avoiding the poison location. Both food and poison locations emit a red light and they can