Lecture 3: Direct Perceiving, Indirect Perceiving (Turvey, 2019, Lectures on Perception)

In this lecture, Turvey provides a formal definition of what it means to claim a theory of perception is direct vs indirect. A theory of direct perception invokes lawfully specifying information, while all indirect theories invoke at least one mediating physiological or psychological process. Direct theories are allowed to discuss internal states, etc (Gibsonian neuroscience isn't a contradiction in terms), but these states are not allowed to alter information so it is no longer lawfully related to the environment. The big Turvey word we will learn about in this chapter is impredicative entailment - exciting! I'll also briefly point to some implications this chapter has for my recent papers with Sabrina on neuroscience and mechanism.

Perceiving, Turvey argues, "...must provide a reliable means or orienting and adjusting activity to the environment." (pg 27). This means perception must trade in activity relevant properties; the objects of perception should be the objects of action. This also implies a mutuality between the organism and the objects of perception/activity; you can only direct an activity (such as standing) towards an environment with physical properties that complement the properties of you involved in standing.

There are two ways to describe a perceptual system that must fill this job description.

1. A direct theory of perception invokes a two-term relation - there is an organism and environment, and they are placed in some relation to one another (e.g. the organism is walking-on the environmental surface). The organism then perceives the environment, and the organism also acts on the environment. 
2. An indirect theory of perception invokes a three-term relation. When the organism is walking on the environmental surface, it's perception requires a mediator or some kind, and what the organism experiences is not the surface, but the mediator. Mediators might be retinal images, etc, but in general these days are some form of representation. 
1. Turvey notes an immediate problem here; it's not clear how to connect indirect perception to action, because the object of perception (the mediator) is not the same as the object of action (the surface) and we do not, for example, walk-on representations of surfaces. 
2. Another issue: even in an indirect theory, the organism always has direct access to something, specifically the final mediator leading to perceptual experience. So what makes a theory a direct perception theory is that the environment can be perceived directly - the two-term relation. 

Information-L is an Impredicative Entailment

In order for us to have unmediated (direct) access to the environment, we need to propose that we interact with something that specifies that environment. Patterns in ambient media that have been lawfully created by properties of the environment can specify those properties; Turvey refers to this as Information-L to distinguish it from Shannon information and to emphasise the role of laws in creating specificity. Specification needs to allow one thing (information) to be about another thing (the environment) without information mediating access to the environment, or else perception of the environment is indirect. To do this, Turvey formalises specification as an impredicative entailment. 

An entailment propagates truth; if X entails Y, then whatever is true of X is true of Y. For X to be about Y, however, you must close the loop; both 'X entails Y' and 'Y entails X' must be true. For example, I can point to a cat and claim that my finger posture is about the cat's location if and only if my finger is pointing to the cat's position AND that cat's position is why my finger has that posture. Closed loop entailments like this are called impredicative. Traditionally, impredicative entailments are considered a problem to be eliminated, because they involve defining elements that make up a system in terms of the system as a whole, which can become a vicious cycle. Turvey will justify us being allowed to rely on them later, and now just notes that complex systems (such as perception-action systems) have inherent impredicative loops. 

Linear Causal Chains vs Nests of Entailments

One reason why impredicativities are generally frowned on is that they involve circular causations; the sub-component is defined in terms of the full system and vice versa. When we think about cause and effect, we tend to analyse things as linear chains of causation - one part of a system causes a state change in the next part, and so on until the system exhibits it's behaviour. These are open, not closed, entailments, and so a behaviour that is accomplished at the end of a chain of events cannot be impredicative.

Another way of thinking about this limitation is that this kind of analysis places the various causes and the resulting effects on the same level of analysis. However, what we name as a 'cause' or a link in the chain is actually a theory-laden hypothesis about what that element is, how it's made and how it does it's work. The label we give this set of 'interlocked conceptual structures' has no causal power, the multi-level elements it names have the power. You can illustrate this with respect to perception. Indirect theories of perception invoke chains of causes and effects that lead from the environment, into the organism and towards the creation of the percept of the environment. The big problem is that the nature of the percept is of a different kind from all the (physical, physiological) links in the causal chain of events. They do not explain why the percept is the way that it is.

Direct theories of perception do not involve causal chains of events, and in line with the Composition, Environment, Structure (CES) model from Lecture 1, perception is not considered to be the end point of a chain of events. Instead, perception-of-E is an accomplishment of the CES system, supported by the nature of the environment, Information-L, and the internal dynamics of the organism that registers Information-L about the environment. We therefore need a different conceptualisation of causation that is up to describing what happens in complex, impredicative CES systems like the perception-action system, specifically nested entailment structures.

Reflections

Why is Information-L not a mediator? In one very real sense, it stands between the organism and the environment and perception does not happen without it. Turvey is arguing that it is not acting as a link in a causal chain of events, rather it is acting like a leg of a table, one of the underlying reasons the table (perception of the environment) is standing up. Perception is not what happens at the end of a chain of events, it is the activity of the system as an impredicative whole. 

I'm very interested in this question, because Sabrina & I have suggested information does work as a mediator in Golonka & Wilson, 2019; we literally call it an ecological representation! That said, we argue very firmly that specification makes it a very different kind of representation from those invoked by indirect theories of perception. But have we made an ontological error? I don't yet think so; I think this analysis (with perhaps some more caution in some of our language) remains viable. But I will leave a detailed analysis of that for a later post, after I've given it more thought.

Impredicativity is also a challenge to our work proposing that the ecological approach can support mechanistic modelling. This kind of work requires you to decompose a system into parts and processes, study those, model them, and then bring them back into a model of the system to explain the behaviour of the system. But impredicativity means the correct parts are defined with respect to the system as a whole, and that kind of circularity seems to imply these systems cannot be meaningfully decomposed. This is Chemero's basic argument, or part of it anyway.

However, I think impredicativity actually helps the argument Sabrina and I were making. We argue that you must identify parts and processes at the scale of tasks; a meaningful part in coordination is not 'the dynamics of human limbs' but 'the dynamics of rhythmically moving human limbs'. This is an impredicative way to define a part, it came very naturally to us as we analysed how Bingham had proceeded, and it seems to work just fine. Again, I'm going to stick a pin in this for later because it deserves a rigorous defence.