1. Predictive coding
As the target article notes, the capacity for knowledge representation is of only limited use in predicting (or in interpreting or explaining) the behavior of others. We can't simply “look at the facts” to predict or explain another's behavior if the other doesn't “share” – that is, doesn't know, isn't aware of – those facts. Nevertheless, it would be folly for the brain to ignore the actual world completely and start over, attempting to build from scratch the set of “facts” that guide the other's behavior. Rather than approach the behavior of others with a blank slate (in Bayesian terms, without priors) – which would be inefficient, if not impossible – the brain very likely implements a predictive strategy. In such a strategy, the actual world – that is, what we ourselves take to be the facts – serves as a starting point, an opening bid or bet, subject to revision (“correction”) on the basis of new evidence (Clark, Reference Clark2013; Koster-Hale & Saxe, Reference Koster-Hale and Saxe2013).
2. Analysis by synthesis
There is, in fact, a plausible mechanism for implementing such a predictive scheme for anticipating and interpreting others' behavior, as I argue in Gordon (Reference Gordon, Ochsner and Gilead2021). It exploits a strategy that appears to operate in several other areas of cognition, including visual and speech perception: that of analysis by synthesis. Specifically, the brain interprets the behavior of others by testing hypothetical ways of generating that behavior. This would involve inverse use of one's own system for planning and generating intentional action: inverse, in that what is “given” is the behavior to be generated, and the “result” is whatever best explains this behavior (Baker, Saxe, & Tenenbaum, Reference Baker, Saxe and Tenenbaum2009; Jara-Ettinger, Reference Jara-Ettinger2019). Such an inversion of the action planning system would run concurrently with its primary “forward” use in generating one's own actions; otherwise, one would have to suspend one's own actions in order to interpret the actions of others. This is consistent with evidence of “motor contagion,” or interference effects between observed and executed actions. There appears to be a competition for neural resources, where the same, or strongly overlapping, resources are employed concurrently in goal-directed action planning and in interpreting the goal-directed actions of others (Blakemore & Frith, Reference Blakemore and Frith2005; Bouquet, Shipley, Capa, & Marshall, Reference Bouquet, Shipley, Capa and Marshall2011).
The inverse use of the planning system for hypothetically generating the actions of others would ordinarily require adjustments of the top-down inputs to the system. These would include adjustments of the factual input, the set of facts that influence planning. In hypothetically generating another's actions, the planning system must be selectively decoupled (disconnected and unplugged) from some of these facts. In the classic “false belief” condition, you see individual A place her treasure at location x. You also see that
(m) the treasure has been moved and is now at a different location y.
If you were planning to steal the treasure, your action planning system would take account of (m) and direct you to location y. However, if your system is hypothetically generating A's plan to retrieve A's treasure, the question arises: Does A know about the move? Is A aware that (m)? The possibility of attributing ignorance, or not knowing, is simply the possibility of decoupling the action planning system from the fact that (m). (Egocentric ignorance acknowledges that there are facts to which our own planning is not yet coupled or connected.) Knowledge, on the contrary, is represented simply by nonintervention. That is, one implicitly attributes knowledge that (m) simply by not decoupling the system from the fact that (m). “Knowledge representations” accordingly consist in nothing more than access to facts.
Attributing ignorance consists of decoupling from fact, which is an extra step beyond implicitly attributing knowledge. False belief requires decoupling as well as introducing into the planning process an “as if” fact, such as that the treasure is still at location x. True belief for the wrong reason would similarly entail introducing an “as if” fact. (Although it might produce the same actions as the “real” fact, the counterfactual dependencies would differ.) The upshot is that what is really basic is a shared world, where, prior to any corrective processing, everything we ourselves regard as the world, as the facts, is publicly accessible and thus available to others as possible reasons for action.
In sum, the commentary presents strong empirical evidence that knowledge is basic; however, I disagree with the explanation. And I heartily approve the “call to arms” at the end. In my own case it's been revelatory to step outside philosophy and consider possible neural mechanisms that might explain, clarify, and validate the intuitive idea that we understand one another as actors in a shared world.
You have
Access
The target article makes a compelling case for an important thesis, the primacy of knowledge attribution. It takes us far in the right direction, veering off only in the final section, where it asks why the capacity for knowledge representation “would have ended up being one that is cognitively basic.” The assumption seems to be that either knowledge attribution or belief attribution might have become the basic one, but a certain important function of knowledge attribution (namely, allowing us to learn about the world from others) caused knowledge attribution to “end up” as basic.
I think knowledge attribution is basic in a much more straightforward way. Some current ideas in cognitive neuroscience – predictive coding and analysis by synthesis – show how factual knowledge may be attributed to others simply by default. Withholding or diminishing that attribution, as in attributing a belief that falls short of knowledge, requires additional steps in neural coding and processing. Those extra steps, their added complexity and their drain on resources, suffice to explain the empirical findings: Why some individuals – nonhuman primates, young children, and certain cognitively impaired people – can attribute knowledge but not belief, whereas none attribute belief but not knowledge; and why attributions of knowledge are “more automatic” than those that require additional processing.
1. Predictive coding
As the target article notes, the capacity for knowledge representation is of only limited use in predicting (or in interpreting or explaining) the behavior of others. We can't simply “look at the facts” to predict or explain another's behavior if the other doesn't “share” – that is, doesn't know, isn't aware of – those facts. Nevertheless, it would be folly for the brain to ignore the actual world completely and start over, attempting to build from scratch the set of “facts” that guide the other's behavior. Rather than approach the behavior of others with a blank slate (in Bayesian terms, without priors) – which would be inefficient, if not impossible – the brain very likely implements a predictive strategy. In such a strategy, the actual world – that is, what we ourselves take to be the facts – serves as a starting point, an opening bid or bet, subject to revision (“correction”) on the basis of new evidence (Clark, Reference Clark2013; Koster-Hale & Saxe, Reference Koster-Hale and Saxe2013).
2. Analysis by synthesis
There is, in fact, a plausible mechanism for implementing such a predictive scheme for anticipating and interpreting others' behavior, as I argue in Gordon (Reference Gordon, Ochsner and Gilead2021). It exploits a strategy that appears to operate in several other areas of cognition, including visual and speech perception: that of analysis by synthesis. Specifically, the brain interprets the behavior of others by testing hypothetical ways of generating that behavior. This would involve inverse use of one's own system for planning and generating intentional action: inverse, in that what is “given” is the behavior to be generated, and the “result” is whatever best explains this behavior (Baker, Saxe, & Tenenbaum, Reference Baker, Saxe and Tenenbaum2009; Jara-Ettinger, Reference Jara-Ettinger2019). Such an inversion of the action planning system would run concurrently with its primary “forward” use in generating one's own actions; otherwise, one would have to suspend one's own actions in order to interpret the actions of others. This is consistent with evidence of “motor contagion,” or interference effects between observed and executed actions. There appears to be a competition for neural resources, where the same, or strongly overlapping, resources are employed concurrently in goal-directed action planning and in interpreting the goal-directed actions of others (Blakemore & Frith, Reference Blakemore and Frith2005; Bouquet, Shipley, Capa, & Marshall, Reference Bouquet, Shipley, Capa and Marshall2011).
The inverse use of the planning system for hypothetically generating the actions of others would ordinarily require adjustments of the top-down inputs to the system. These would include adjustments of the factual input, the set of facts that influence planning. In hypothetically generating another's actions, the planning system must be selectively decoupled (disconnected and unplugged) from some of these facts. In the classic “false belief” condition, you see individual A place her treasure at location x. You also see that
(m) the treasure has been moved and is now at a different location y.
If you were planning to steal the treasure, your action planning system would take account of (m) and direct you to location y. However, if your system is hypothetically generating A's plan to retrieve A's treasure, the question arises: Does A know about the move? Is A aware that (m)? The possibility of attributing ignorance, or not knowing, is simply the possibility of decoupling the action planning system from the fact that (m). (Egocentric ignorance acknowledges that there are facts to which our own planning is not yet coupled or connected.) Knowledge, on the contrary, is represented simply by nonintervention. That is, one implicitly attributes knowledge that (m) simply by not decoupling the system from the fact that (m). “Knowledge representations” accordingly consist in nothing more than access to facts.
Attributing ignorance consists of decoupling from fact, which is an extra step beyond implicitly attributing knowledge. False belief requires decoupling as well as introducing into the planning process an “as if” fact, such as that the treasure is still at location x. True belief for the wrong reason would similarly entail introducing an “as if” fact. (Although it might produce the same actions as the “real” fact, the counterfactual dependencies would differ.) The upshot is that what is really basic is a shared world, where, prior to any corrective processing, everything we ourselves regard as the world, as the facts, is publicly accessible and thus available to others as possible reasons for action.
In sum, the commentary presents strong empirical evidence that knowledge is basic; however, I disagree with the explanation. And I heartily approve the “call to arms” at the end. In my own case it's been revelatory to step outside philosophy and consider possible neural mechanisms that might explain, clarify, and validate the intuitive idea that we understand one another as actors in a shared world.
Financial support
This research received no specific grant from any funding agency, commercial, or not-for-profit sectors.
Conflict of interest
None.