Active inference, causal cognition, and structure in the world

In the most general terms, causal cognition could be defined as information processing that involves, more or less explicitly, an appreciation of a notion of causality on the part of an intelligent agent. To a first approximation, appreciating a notion of causality might involve any combination of abilities such as the understanding, exploitation, discovery, and construction of cause-effect relationships (among other things). An agent with that kind of sensitivity could be called a causal agent.

Active inference is a theoretical framework developed by Karl Friston and colleagues over many years to account for the computational function(s) of the nervous system and, more fundamentally, self-organization and sentience in complex systems (Friston, 2005, 2019; Friston & Kiebel, 2009; Parr et al., 2020, 2022). Several mathematical formulations of the framework have been used to explain neurocognitive activity in different domains and applied to model neural and behavioural data in computational cognitive neuroscience (Adams et al., 2013; Buckley et al., 2017; Da Costa et al., 2020; Friston, FitzGerald, Rigoli, Schwartenbeck, & Pezzulo, 2017; Friston et al., 2010; Mirza et al., 2016, 2019; Parr & Friston, 2017; Pezzulo et al., 2015, 2018; Seth & Friston, 2016). Also, it has received a lot of philosophical attention and the core ideas have been popularised with the names of predictive processing (PP) and prediction error minimization (PEM) (Clark, 2013, 2016; Hohwy, 2013, 2020; Wiese & Metzinger, 2017).

The fundamental and overarching idea of active inference is that information processing in the brain can be seen as a manifestation of some kind of pervasive predictive activity that approximates (hierarchical) Bayesian inference (Friston, 2008; Friston, FitzGerald, Rigoli, Schwartenbeck, & Pezzulo, 2017; Friston, Parr, & de Vries, 2017; Lee & Mumford, 2003). The brain is thought to implement and continuously refine a (hierarchical) generative model of sensory information (Kiefer & Hohwy, 2018). In active inference circles, it is ofthen claimed that an agent equipped with such a generative model is capable of inferring the causes of sensory signal and/or learn about causal structure in the world (see, e.g., Clark, 2016, p. 171; Friston, 2012, p. 2101; Hohwy, 2013, p. 228). Claims like the above may persuasively suggest that active inference agent are causal agent in the sense introduced at the outset; in other words, the overarching theoretical power of active inference would also throw light on causal cognition.

The main goal of the dissertation is to look deep into those claims and ascertain whether they stand up to scrutiny. This will be accomplished in two different ways. First, research from the field of causal machine learning will provide important theoretical coordinates against which to clarify notions like causes and causal structure, and give a sense of how (machine) learning methods can be deployed to discover them from observations (Burgess et al., 2018; Higgins, Matthey, et al., 2017; Schölkopf & von Kügelgen, 2022; Schölkopf et al., 2012, 2021). Second, research on causal cognition in non-human and human animals will form a bedrock of behavioural data on which to base experimentally informed distinctions about different levels or dimensions of causal cognition (Penn & Povinelli, 2007; Starzak & Gray, 2021; Visalberghi & Tomasello, 1998; Woodward, 2011). By combining these different (but related) approaches to dissect active inference as far as causal cognition is concerned, the outcome will be a more refined understanding of the extent to which active inference agents can be regarded as causal agents.