Language and the symbolic use of labels underpin human cognition. As these features imply back and forths between label and labelled object, they require relational symmetry: the capacity to reverse a learned relation A->B into B->A. In non-human animals, such flexible encoding of bidirectional relations has been difficult to obtain experimentally, mostly due to their failure to reverse stimulus relations. This apparent absence of symmetry suggests an evolutionary gap at the origin of human language. Yet, a number of procedural biases, inlcuding perceptual ones, may account for this failure to demonstrate symmetry. Moreover, all studies have relied on specific motor outputs to reveal symmetry, whereas symmetric relations may simply not be learned in such an explicit form. Using the baboon as model primate species, and a worldwide unique, automatized, high-throughput behavioral platform as experimental system, we will conduct new experiments to re-assess this question. We will implement relational learning procedures which correct for all biases identified in past experiments, and rely on implicit rather than explicit measurements to test for the emergence of symmetry. Furthermore, we will study the effects of symmetry on stimulus networks, to assess the baboons’ abilities for flexible relational encoding when more stimuli are involved. A crucial feature of natural intelligence such as relational symnetry ought to be also considered in Artificial Intelligence (AI) systems, as their relational learning abilities are still limited. We will thus investigate the capacity of existing connectionnist models (e.g. Recurrent Neural Networks) to learn symmetrical associations of stimuli, and will explore new architectures able to implement symmetry. Both symmetry or its absence in baboons would inform about the fundaments of human cognition. Hence, far-reaching implications are expected from this project, which may open new avenues of research in both AI and psychology.
Call for proposal
See other projects for this call