Which structures are out there? : Learning predictive compositional concepts based on social sensorimotor explorations

Abstract

How do we learn to think about our world in a flexible, compositional manner? What is the actual content of a particular thought? How do we become language ready? I argue that free energy-based inference processes, which determine the learning of predictive encodings, need to incorporate additional structural learning biases that reflect those structures of our world that are behaviorally relevant for us. In particular, I argue that the inference processes and thus the resulting predictive encodings should enable (i) the distinction of space from entities, with their perceptually and behaviorally relevant properties, (ii) the flexible, temporary activation of relative spatial relations between different entities, (iii) the dynamic adaptation of the involved, distinct encodings while executing, observing, or imagining particular interactions, and (iv) the development of a – probably motor-grounded – concept of forces, which predictively encodes the results of relative spatial and property manipulations dynamically over time. Furthermore, seeing that entity interactions typically have a beginning and an end, free energy-based inference should be additionally biased towards the segmentation of continuous sensorimotor interactions and sensory experiences into events and event boundaries. Thereby, events may be characterized by particular sets of active predictive encodings. Event boundaries, on the other hand, identify those situational aspects that are critical for the commencement or the termination of a particular event, such as the establishment of object contact and contact release. I argue that the development of predictive event encodings naturally lead to the development of conceptual encodings and the possibility of composing these encodings in a highly flexible, semantic manner. Behavior is generated by means of active inference. The addition of internal motivations in the form of homeostatic variables focusses our behavior – including attention and thought – on those environmental interactions that are motivationally-relevant, thus continuously striving for internal homeostasis in a goal-directed manner. As a consequence, behavior focusses cognitive development towards (believed) bodily and cognitively (including socially) relevant aspects. The capacity to integrate tools and other humans into our minds, as well as the motivation to flexibly interact with them, seem to open up the possibility of assigning roles – such as actors, instruments, and recipients – when observing, executing, or imagining particular environmental interactions. Moreover, in conjunction with predictive event encodings, this tool- and socially-oriented mental flexibilization fosters perspective taking, reasoning, and other forms of mentalizing. Finally, I discuss how these structures and mechanisms are exactly those that seem necessary to make our minds language ready.