The construction of concepts: Active inference and the probabilistic language of thought

The active inference framework is increasingly influential as a unifying model of perception, action, and cognition. However, it is unclear how higher cognition, thought, and concept formation should be understood within active inference. In this paper, we argue that active inference has the resources to deliver an account of the Language of Thought. The Language of Thought Hypothesis (LoTH) posits that cognitive processes are underpinned by a symbolic system of internal representations akin to a language (Fodor, 2010). This "mentalese," as it is sometimes referred to, supports the notions of systematicity and productivity, suggesting that thinking operates via syntactic rules and structured symbols, allowing for the capacity to produce and understand novel cognitive constructs (Fodor, 1998). While LoTH fell out of favour, it has had a recent resurgence, and a particularly promising version of it is the probabilistic Language of Thought (pLoT) which adds probabilistic processes to the LoT picture. We argue that the fundamental features of conceptual representation posited by the pLoT can be accounted for in active inference when we consider concepts as built up sequentially through a series of mental actions, or mental action policies articulated in active inference. Once constructed, concepts can be compressed, are then amenable to rapid deployment and can serve as primitives in the construction of more complex concepts.