A Dual-Process Theory of Mindreading

Click here and press the right key for the next slide (or swipe left)

also ...

Press the left key to go backwards (or swipe right)

Press n to toggle whether notes are shown (or add '?notes' to the url before the #)

Press m or double tap to slide thumbnails (menu)

Press ? at any time to show the keyboard shortcuts

A Dual-Process Theory of Mindreading

Dual Process Theory of Mindreading (core part)

Two (or more) mindreading processes are distinct:
the conditions which influence whether they occur,
and which outputs they generate,
do not completely overlap.

What models of minds and actions, and of behaviours,

and what kinds of processes,

underpin mental state tracking in different animals?

Way forward:

1. Construct a theory of behaviour reading

2. Construct a theory of mindreading

This is what we are in the process of doing ... That there is more to be done is suggested by the three questions about mindreading, only one of which we can so far answer. What is still to be done?

Recall David Marr’s famous three-fold distinction between levels of description of a system: the computational theory, the

This is easy to understand in simple cases. To illustrate, consider a GPS locator. It receives information from four satellites and tells you where on Earth the device is.

There are three ways in which we can characterise this device.

1. computational description

First, we can explain how in theory it is possible to infer the device’s location from it receives from satellites. This involves a bit of maths: given time signals from four different satellites, you can work out what time it is and how far you are away from each of the satellites. Then, if you know where the satellites are and what shape the Earth is, you can work out where on Earth you are.

-- What is the thing for and how does it achieve this?

The computational description tells us what the GPS locator does and what it is for. It also establishes the theoretical possibility of a GPS locator.

But merely having the computational description does not enable you to build a GPS locator, nor to understand how a particular GPS locator works. For that you also need to identify representations and alogrithms ...

2. representations and algorithms

At the level of representations and algorthms we specify how the GPS receiver represents the information it receives from the satellites (for example, it might in principle be a number, a vector or a time). We also specify the algorithm the device uses to compute the time and its location. The algorithm will be different from the computational theory: it is a procedure for discovering time and location. The algorithm may involve all kinds of shortcuts and approximations. And, unlike the computational theory, constraints on time, memory and other limited resources will be evident.

So an account of the representations and algorithms tells us ...

-- How are the inputs and outputs represented, and how is the transformation accomplished?

3. hardware implementation

The final thing we need to understand the GPS locator is a description of the hardware in which the algorithm is implemented. It’s only here that we discover whether the device is narrowly mechanical device, using cogs, say, or an electronic device, or some new kind of biological entity.

-- How are the representations and algorithms physically realised?

The hardware implementation tells us how the representations and algorithms are represented physically.

Marr (1992, 22ff)

Where does minimal theory of mind fit in here?

This is a description of a model of mind at the level of a computational theory; it is completely agnostic about representations and algorithms.

\citet[p.~22ff]{Marr:1982kx} distinguishes:

\begin{itemize}

\item computational description---What is the thing for and how does it achieve this?

\item representations and algorithms---How are the inputs and outputs represented, and how is the transformation accomplished?

\item hardware implementation---How are the representations and algorithms physically realised?

\end{itemize}

Recall David Marr’s famous three-fold distinction between levels of description of a system: the computational theory, the

This is easy to understand in simple cases. To illustrate, consider a GPS locator. It receives information from four satellites and tells you where on Earth the device is.

There are three ways in which we can characterise this device.

1. computational description

-- What is the thing for and how does it achieve this?

The computational description tells us what the GPS locator does and what it is for. It also establishes the theoretical possibility of a GPS locator.

2. representations and algorithms

So an account of the representations and algorithms tells us ...

-- How are the inputs and outputs represented, and how is the transformation accomplished?

3. hardware implementation

-- How are the representations and algorithms physically realised?

The hardware implementation tells us how the representations and algorithms are represented physically.

Marr (1992, 22ff)

What we need to make progress with the three questions is a hypothesis about the processes (representations and algorithms in Marr’s terms).

Why is belief-tracking in adults sometimes but not always automatic?

How could belief-tracking ever be automatic given evidence that it depends on working memory and consumes attention?

Why are there dissociations in nonhuman apes’ performance on belief-tracking tasks?

To answer these questions, and to construct a theory of mindreading, we need to shift to a hypothesis about the processes (representations and algorithms in Marr’s terms).

a dual-process theory

Start with a simple causal model.

‘response 1’ is a variable representing which response the subject will give. [Which values it takes will depend on what sort of response it is (e.g. a verbal response, proactive gaze, button press.) We can think of it as taking three values, one for correct belief tracking, one for fact tracking, and one for any other response.]

‘process 1’ and ‘process 2’ are variables which each represent whether a certain kind of mindreading process will occur and, if so, what it’s outcome is.

And the arrows show that the probability that response 1 will have a certain value is influenced by the value of the variables process 1 and process 2 (and by other things not included in the model). So it should be possible to intervene on the value of ‘process 1’ in order to bring about a change in the value of ‘response 1’.

[I’ve used thicker and thinner arrows informally to indicate stronger and weaker dependence. Strictly speaking the width has no meaning and this model doesn’t specify exactly how the values of variables are related, only that they are.]

Of course, much the same is true for ‘response 2’ as well. It’s just that the changing the values of other variables will have different effects on the values of ‘response 1’ and ‘response 2’.

[Limit: this depiction ignores time, which is of course critical.]

We must avoid a false assumption about the relation between types of response and kinds of process ...

Process 1 -> Response 1

Process 2 -> Response 2

The false assumption is that responses of type R1 are dominated by one mindreading process whereas responses of type R2 are dominated by another mindreading process. But responses types and processes may not be so closely associated, of course.

This is because (a) any response is likely to be a consequence of multiple processes; and, (b), for some response types such as button selection or proactive gaze, changing factors like time pressure could change which mindreading process dominates responses of that type.% \footnote{% In fact variability in the relation between a mindreading process and a response type is a potentially useful source of evidence in support of a dual process theory of mindreading. Changes in the processes determining a response type can be detected where we have situations in which we know, or assume, the two processes yield different answers (cf Todd et al). }

So what does a Dual Process Theory of Mindreading claim? The core claim is just this:

Dual Process Theory of Mindreading (core part)

Two (or more) mindreading processes are distinct:
the conditions which influence whether they occur,
and which outputs they generate,
do not completely overlap.

\textbf{You might say, this is a schematic claim, one totally lacking substance.} You’d be right: and that’s exactly the point.

A key feature of this Dual Process Theory of Mindreading is its \textbf{theoretical modesty}: it involves no a priori committments concerning the particular characteristics of the processes.

1. Mechanisms

Conjecture: Some mindreading processes are more automatic than others.

Prediction: Varying task instructions will differentially affect responses indicative of mindreading.

2. Models

Conjecture: Some automatic mindreading processes rely on minimal models of the mental.

Prediction: Those mindreading processes are subject to the signature limits of minimal models.

(See \citealp{wang:2015_limits,low:2010_preschoolers,low:2014_quack,edwards:2017_reaction}; contrast \citealp{scott:2015_infants}.)

Conjecture: Some automatic mindreading depends on motor processes.

Prediction: Impairing motor processes can impair mindreading.

evidence?

Is there any evidence for the dual-process theory?

Take a pair of response types, R1 and R2, which can both indicate belief tracking (or, thinking more generally, perspective taking) under some conditions. For example, consider initial proactive gaze (R1) contrasted with final button selection (R2).

The conditions under which one or another response type tracks beliefs varies ...

1.a E.g. changing the \textbf{instructions} reduces the probability that responses of one type will track beliefs without much affecting the probability that responses of the other type will track belief.

Direct evidence for this includes studies by Schneider et al. (Would be good to have more given some difficulties in replicating these; although note that Schneider et al previously replicated these results several times).

Indirect evidence for this involving responses such as RTs and motion trajectories comes from Edwards and Low, 2017; Kovacs et al; van der Wel et el.

1.b Or, e.g., changing the \textbf{content of the belief} to be tracked from location to identity, say, reduces the probability that responses of one type track beliefs without much affecting the probability that responses of the other type will track beliefs.

Evidence for this includes studies by Low and his collaborators.

So it should be possible find different conditions that reduce the probability that different responses track beliefs.

Further, the probability that responses of these two types track beliefs should be differently affected by factors such as time pressure and cognitive load. (It isn’t important that either is entirely unaffected; what matters is just that the effects are different.)

As far as I know we don’t yet have direct evidence for this, because few studies have compared what happens to two response types as factors like time pressure or cognitive load are varies.

There is, of course, some indirect evidence for this from studies which show that responses of some types are less susceptible to cognitive load than others (e.g. \citet{qureshi:2010_executive} on L1 VPT).

So if we think just about adult humans’ mindreading (or chimpanzee mindreading, for that matter), it turns out that when we consider different ways of measuring mindreading, we find a pattern of dissociations between performance.

(a) Automatic processes useless if they only control proactive gaze (b) Infants can demonstrate belief tracking not only in v-of-e and anticipatory looking, but also on tasks which require actions such as pointing and helping.

Can a dual process theory of mindreading accommodate this objection?

This question came up in discussion with Jason Low and together with Chris Maymon we found a way to answer it. Our conjecture was that athough automatic mindreading processes are unlikely to feed into deliberative reasoning about what to do, they could affect motor actions and so initiate action in the right direction.

To test this conjecture, we developed a simple helping paradigm.

\section{Maymon, Sivanantham, Low & Butterfill (pilot)} \begin{center} \includegraphics[scale=0.37]{img/maymon_fig1.png} \end{center} Sequence of events (1 – 10) in the FB-identity condition. \begin{center} \includegraphics[scale=0.3]{img/maymon_fig3.png} \end{center} \% of participants displaying type of first look by condition (**P = 0.005; ***P < 0.001). \vfill \begin{center} \includegraphics[scale=0.35]{img/maymon_fig4.png} \end{center} Schematic representation of individuals’ (N = 96) course of action in Experiment 1 between conditions: (A) FB-identity, (B) FB-location, (C) TB-identity, and (D) TB-location. The course was divided into 4 stages: (1) swerving, (2) advancing, (3) reaching, and (4) ultimately handing over the actual or non-actual bag (dotted lines represent thresholds for each stage). In Experiment 2, we examined how stalling of motor representations, by temporarily tying individual observers’ hands (E), affected the course of their (N = 24) helping action in the FB-location condition (F).

conjecture

although representations processed automatically may not feature in practical reasoning,

automatic mindreading could affect motor actions

and so initiate action in the right direction.

Maymon et al, pilot Figure 1, used with permission

Maymon et al, pilot Figure 1, used with permission

Participants started while seated at a desk; they needed to get up to help a protagonist by taking their chair, standing on it and retrieving a bag from a high place.

Maymon et al (pilot), figure 2, used with permission

First fixations show just the pattern you would expect for beliefs about locations.

Maymon et al (pilot), figure 2, used with permission

First fixations also show just the pattern our dual process theory predicts for beliefs about identity.

Maymon et al (pilot), figure 2, used with permission

Maymon et al divided participants’ movements into four stages: swerving (stage-1), advancing (stage-2) towards the actual or non-actual bag, reaching (stage-3) for the bag, and ultimately handing over (stage-4) the actual or non-actual bag.

The first stage of movement shows just the pattern that first fixation did, and indeed in FB-LOC, first fixation and swerve are correlated.

Maymon et al (pilot), figure 4B, used with permission

Here you can see how participants moved in the false belief location condition. There was also a true belief condition.

As you can see, participants mostly started off in the right direction. This is the swerve.

But there is quite a bit of switching, and a number of

The swerve is correlated with the protagonists’ belief state: which direction they swerve in differs between true and false beliefs. The direction of swerve is also correlated with proactive gaze

How could implicit mindreading influence action
independently of practical reasoning?

Or, to put the question differently, How does tracking someone’s beliefs enable humans to be better cooperators when helping them with a task?

We conjectured that it might do so by influencing motor representations and processes. To test this conjecture, we re-ran our helping study, but this time we had the subjects hands bound behind their backs ...

Maymon et al (pilot), figure 4E-F, used with permission

Hands tied (a) eliminates anticipatory looking indicative of belief tracking, and (b) destorys the swerve indicative of belief tracking.

NB: we compared hands-tied with hands-free FB location, and separately with hands-free TB location. We did not do hands-tied TB. (‘proactive first looks and swerving of subjects in the hands-tied FB-location condition resembled proactive first looks and swerving of hands-free subjects in the corresponding TB-location condition (P = 0.566 and P = 0.057, respectively). ’)

[Stress that the new conjecture, automatic mindreading is an extension of motor processes that underpin goal tracking, is an optional development of the dual process theory and not something entailed by it.]

So our results (which are very preliminary and may not be replicated--we will find out in a few weeks) support the view that \textbf{to better cooperate with others, what others believe must influence how we represent actions motorically}.

Implicit mindreading doesn’t control how you end up acting, but it can set you off in the right direction.

Prediction:

Slowing 1- or 2-year-olds’ helping responses would reduce the probability that their responses indicate belief tracking.

(And the opposite may be true for adults.)

So what does a Dual Process Theory of Mindreading claim? The core claim is just this:

Dual Process Theory of Mindreading (core part)

Two (or more) mindreading processes are distinct:
the conditions which influence whether they occur,
and which outputs they generate,
do not completely overlap.

\textbf{You might say, this is a schematic claim, one totally lacking substance.} You’d be right: and that’s exactly the point.

A key feature of this Dual Process Theory of Mindreading is its \textbf{theoretical modesty}: it involves no a priori committments concerning the particular characteristics of the processes.

1. Mechanisms

Conjecture: Some mindreading processes are more automatic than others.

Prediction: Varying task instructions will differentially affect responses indicative of mindreading.

2. Models

Conjecture: Some automatic mindreading processes rely on minimal models of the mental.

Prediction: Those mindreading processes are subject to the signature limits of minimal models.

Conjecture: Some automatic mindreading depends on motor processes.

Prediction: Impairing motor processes can impair mindreading.

What about the three questions?

So the modest dual-process theory I have been advocating can generate readily testable predictions, and there is some evidence for it. But does it also help to answer the three questions about mindreading?

Why is belief-tracking in adults sometimes but not always automatic?

How could belief-tracking ever be automatic given evidence that it depends on working memory and consumes attention?

Why are there dissociations in nonhuman apes’ performance on belief-tracking tasks?

Keyboard Shortcuts?

A Dual-Process Theory of Mindreading

Keyboard Shortcuts`?`