Enactivism, Computation, and Autonomy

Enactivism has historically rejected computational characterisations of cognition, at least in more traditional versions. This has led to the perception that enactivist approaches to cognition must be opposed to be more mainstream computationalist approaches, which offer a computational characterisation of cognition.

However, the conception of computation which enactivism rejects is in some senses quite old fashioned, and it is not so clear that enactivism need necessarily be opposed to computation, understood in a more modern sense. Demonstrating that there could be compatibility, or at least not a necessary opposition, between enactivism and computationalism (in some sense) would open the door to a possible reconciliation or cooperation between the two approaches.

In a recently published paper (Villalobos & Dewhurst 2017), my collaborator Mario and I have focused on elucidating some of the reasons why enactivism has rejected computation, and have argued that these do not necessarily apply to more modern accounts of computation. In particular, we have demonstrated that a physically instantiated Turing machine, which we take to be a paradigmatic example of a computational system, can meet the autonomy requirements that enactivism uses to characterise cognitive systems. 

This demonstration goes some way towards establishing that enactivism need not be opposed to computational characterisations of cognition, although there may be other reasons for this opposition, distinct from the autonomy requirements.

The enactive concept of autonomy first appears in its modern guise in Varela, Thompson, & Rosch’s 1991 book The Embodied Mind, although it has important historical precursors in Maturana’s autopoietic theory (see his 1970, 1975, 1981; see also Maturana & Varela 1980) and cybernetic work on homeostasis (see eg Ashby 1956, 1960). There are three dimensions of autonomy that we consider in our analysis of computation. 

Self-determination requires that the behaviour of an autonomous system must be determined by that system’s own structure, and not by external instruction. 

Operational closure requires that the functional organisation of an autonomous system must loop back on itself, such that the system possesses no (non-arbitrary) inputs or outputs. 

Finally, an autonomous system must be precarious, such that the continued existence of the system depends on its own functional organisation, rather than on external factors outside of its control. In this post I will focus on demonstrating that these criteria can be applied to a physical computing system, rather than addressing why or how enactivism argues for them in the first place.

All three criteria have traditionally been used to disqualify computational systems from being autonomous systems, and hence to deny that cognition (which for enactivists requires autonomy) can be computational (see eg Thompson 2007: chapter 3). Here it is important to recognise that the enactivists have a particular account of computation in mind, one that they have inherited from traditional computationalists.

According to this ‘semantic’ account, a physical computer is defined as a system that performs systematic transformations over content-bearing (ie representational) states or symbols (see eg Sprevak 2010). With such an account in mind, it is easy to see why the autonomy criteria might rule out computational systems. We typically think of such a system as consuming symbolic inputs, which it transforms according to programmed instructions, before producing further symbolic outputs.

Already this system has failed to meet the self-determination and operational closure criteria. Furthermore, as artefactual computers are typically reliant on their creators for maintenance, etc., they also fail to meet the precariousness criteria. So, given this quite traditional understanding of computation, it is easy to see why enactivists have typically denied that computational systems can be autonomous.

Nonetheless, understood according to more recent, ‘mechanistic’ accounts of computation, there is no reason to think that the autonomy criteria must necessarily exclude computational systems. Whilst they differ in some details, all of these accounts deny that computation is inherently semantic, and instead define physical computation in terms of mechanistic structures. 

We will not rehearse these accounts in any detail here, but the basic idea is that physical computation can be understood in terms of mechanisms that perform systematic transformations over states that do not possess any intrinsic semantic content (see eg Miłkowski 2013; Fresco 2014; Piccinini 2015). With this rough framework in mind, we can return to the autonomy criteria.

Even under the mechanistic account, computation is usually understood in terms of mappings between inputs and outputs, where there is a clear sense of the beginning and end of the computational operation. A system organised in this way can be described as ‘functionally open’, meaning that its functional organisation is open to the world. 

A functionally closed system, on the other hand, is one whose functional organisation loops back through the world, such that the environmental impact of the system’s ‘outputs’ contributes to the ‘inputs’ that it receives. A simple example of this distinction can be found by considering two different ways that a thermostat could be used. 

In the first case the sensor, which detects ambient temperature, is placed in one house, and the effector, which controls a radiator, is placed in another (see figure 1). This system is functionally open, because there is only a one-way connection between the sensor and the effector, allowing us to straightforwardly identify inputs and outputs to the system.