The target article develops a theory of abstraction largely focused on representations of objects and states of the external world. It argues that predictive cognition needs an account of hierarchical representational diversity, partly to explain the human capacity to transcend the “here and now.” So far, so good.
Although the topic is not explicitly addressed, the arguments concerning degrees of abstraction, and the case for a hierarchy, apply just as forcefully in the case of representations of the values of, or predicted returns from, behaviours and actions as they do to representations of the external world. (Gilead et al. say that their theory is intended to apply to “mental states” including desires and intentions, but go on to focus almost exclusively on beliefs.) Evolved cognition, after all, and whatever you think of predictive processing, isn't a goal in itself, but a means to co-ordinating behaviour with contingencies and changes in the body and world. The actions available to an agent typically have varying and multi-modal costs and returns. The costs include time, energy, depletion of specific resources such as water and salt, exposure to various risks, and the opportunity cost of forgone actions. The returns can include hydration, the many dimensions of nutrition, rest, safety, territory, and opportunities to mate. There are, that is, many concrete dimensions of cost and benefit.
At the extreme of low abstraction, specific actions can be occasioned by detections of specific concrete, single modality returns or costs (e.g., dehydration, or estimated effort). At the other, action can be selected and prioritised on the basis of wholly abstract considerations, such as utility. In evolutionary settings, utility can be taken to correspond to fitness, permitting economic and decision-theoretic analyses of rational agency to be applied to evolutionary analyses of optimality. Okasha (Reference Okasha2013), for example, argues that an agent that acts and chooses as if performing Bayesian updating is rationally optimal, and hence a plausible theoretical target for the influence of natural selection on its dispositions and their cognitive implementation. Behavioural ecology is methodologically committed – even if only as a regulative principle – to the determinacy of the fitness return of all behaviour, a view expressed in McNamara and Houston's assertion that any “attempt to understand behavior in terms of the evolutionary advantage that it might confer has to find a ‘common currency’ for comparing the costs and benefits of various alternative courses of action” (McNamara & Houston Reference McNamara and Houston1986, p. 358).
Cognitively useable utility representations would need to integrate the various dimensions of cost and return, in ways sensitive to relations of substitutability between the dimensions and the ways they combine in external objects. (Individual food items, e.g., combine various nutrients, the current value of each of which will depend on the state of the body, and the varying costs of getting access to the food.) Utility representations suitable for guiding action would need to do the same for the various capacities of the body itself, because actions are more or less substitutable with each other, and some goals achievable by more or less large collections of different deployments of the powers of the body. Arguably, some efficiencies in allocation can only be achieved by relatively abstract value representations (Spurrett Reference Spurrett2019).
It isn't clear whether our utility representations are fully abstract. Optimists on this question, especially some neuroeconomists, think there's good evidence that humans and some non-human animals process highly abstract utility representations for rewards in widely varying modalities. Levy and Glimcher (Reference Levy and Glimcher2012), for example, survey studies finding consistent neural signatures of the value of monetary gains and losses, cumulative monetary rewards, anticipation of varying monetary rewards, expected values of uncertain monetary rewards, and discounted value of delayed monetary rewards. More tellingly, they review studies involving choices with at least one incentive other than money, including consumer goods, gustatory rewards (water, juice, and food), physical pain, and social reputation, still finding the same general neural signature. Those who are less optimistic point to the ways in which learning about rewards exhibits failures of abstraction. Rolls (Reference Rolls2013), for example, argues that Levy and Glimcher's evidence doesn't decide between a situation where there is a genuine “common currency” (fully abstract utility) and one where different rewards compete on a “common scale” whereas retaining important differences in their links to specific rewards and courses of action. We've long known, furthermore, that the form of conditioned behaviours is related to the reward they deliver, for example that pigeons peck a key leading to water delivery with a drinking-appropriate action, whereas a key leading to food elicits an eating peck (Jenkins & Moore Reference Jenkins and Moore1973).
We don't need to settle the question of whether human value representations are fully abstract here. What matters is that abstraction applies to costs and returns, and our value representations are abstract to a fairly high degree. This is, furthermore, crucial for useful mental time travel (transcending the “here and now”). If time-travel is to pay its way, as it must have for selective processes to have designed it, it must have contributed to the general function of cognition. That is to say, agents transcend the here and now in order better to determine what to do, here and now. Prospection without reasonably accurate estimation of the costs and returns from the implied courses of action, including extended chains of them, is frivolous speculation.
The neglect of these issues in a target article with “predictive brain” in the title is striking partly because it was in neuroeconomics that the first compelling evidence of prediction-error-based processing in natural brains was found. Even if, in the theoretical limit, reference to reward is replaced by talk of “hyper priors” or some other term from the predictive processing framework, it's still important to recognise and take account of the ways in which abstraction is significant for reward based or evaluative representations, and how such representations are important for overall organismic efficiency.
You have
Access
The target article develops a theory of abstraction largely focused on representations of objects and states of the external world. It argues that predictive cognition needs an account of hierarchical representational diversity, partly to explain the human capacity to transcend the “here and now.” So far, so good.
Although the topic is not explicitly addressed, the arguments concerning degrees of abstraction, and the case for a hierarchy, apply just as forcefully in the case of representations of the values of, or predicted returns from, behaviours and actions as they do to representations of the external world. (Gilead et al. say that their theory is intended to apply to “mental states” including desires and intentions, but go on to focus almost exclusively on beliefs.) Evolved cognition, after all, and whatever you think of predictive processing, isn't a goal in itself, but a means to co-ordinating behaviour with contingencies and changes in the body and world. The actions available to an agent typically have varying and multi-modal costs and returns. The costs include time, energy, depletion of specific resources such as water and salt, exposure to various risks, and the opportunity cost of forgone actions. The returns can include hydration, the many dimensions of nutrition, rest, safety, territory, and opportunities to mate. There are, that is, many concrete dimensions of cost and benefit.
At the extreme of low abstraction, specific actions can be occasioned by detections of specific concrete, single modality returns or costs (e.g., dehydration, or estimated effort). At the other, action can be selected and prioritised on the basis of wholly abstract considerations, such as utility. In evolutionary settings, utility can be taken to correspond to fitness, permitting economic and decision-theoretic analyses of rational agency to be applied to evolutionary analyses of optimality. Okasha (Reference Okasha2013), for example, argues that an agent that acts and chooses as if performing Bayesian updating is rationally optimal, and hence a plausible theoretical target for the influence of natural selection on its dispositions and their cognitive implementation. Behavioural ecology is methodologically committed – even if only as a regulative principle – to the determinacy of the fitness return of all behaviour, a view expressed in McNamara and Houston's assertion that any “attempt to understand behavior in terms of the evolutionary advantage that it might confer has to find a ‘common currency’ for comparing the costs and benefits of various alternative courses of action” (McNamara & Houston Reference McNamara and Houston1986, p. 358).
Cognitively useable utility representations would need to integrate the various dimensions of cost and return, in ways sensitive to relations of substitutability between the dimensions and the ways they combine in external objects. (Individual food items, e.g., combine various nutrients, the current value of each of which will depend on the state of the body, and the varying costs of getting access to the food.) Utility representations suitable for guiding action would need to do the same for the various capacities of the body itself, because actions are more or less substitutable with each other, and some goals achievable by more or less large collections of different deployments of the powers of the body. Arguably, some efficiencies in allocation can only be achieved by relatively abstract value representations (Spurrett Reference Spurrett2019).
It isn't clear whether our utility representations are fully abstract. Optimists on this question, especially some neuroeconomists, think there's good evidence that humans and some non-human animals process highly abstract utility representations for rewards in widely varying modalities. Levy and Glimcher (Reference Levy and Glimcher2012), for example, survey studies finding consistent neural signatures of the value of monetary gains and losses, cumulative monetary rewards, anticipation of varying monetary rewards, expected values of uncertain monetary rewards, and discounted value of delayed monetary rewards. More tellingly, they review studies involving choices with at least one incentive other than money, including consumer goods, gustatory rewards (water, juice, and food), physical pain, and social reputation, still finding the same general neural signature. Those who are less optimistic point to the ways in which learning about rewards exhibits failures of abstraction. Rolls (Reference Rolls2013), for example, argues that Levy and Glimcher's evidence doesn't decide between a situation where there is a genuine “common currency” (fully abstract utility) and one where different rewards compete on a “common scale” whereas retaining important differences in their links to specific rewards and courses of action. We've long known, furthermore, that the form of conditioned behaviours is related to the reward they deliver, for example that pigeons peck a key leading to water delivery with a drinking-appropriate action, whereas a key leading to food elicits an eating peck (Jenkins & Moore Reference Jenkins and Moore1973).
We don't need to settle the question of whether human value representations are fully abstract here. What matters is that abstraction applies to costs and returns, and our value representations are abstract to a fairly high degree. This is, furthermore, crucial for useful mental time travel (transcending the “here and now”). If time-travel is to pay its way, as it must have for selective processes to have designed it, it must have contributed to the general function of cognition. That is to say, agents transcend the here and now in order better to determine what to do, here and now. Prospection without reasonably accurate estimation of the costs and returns from the implied courses of action, including extended chains of them, is frivolous speculation.
The neglect of these issues in a target article with “predictive brain” in the title is striking partly because it was in neuroeconomics that the first compelling evidence of prediction-error-based processing in natural brains was found. Even if, in the theoretical limit, reference to reward is replaced by talk of “hyper priors” or some other term from the predictive processing framework, it's still important to recognise and take account of the ways in which abstraction is significant for reward based or evaluative representations, and how such representations are important for overall organismic efficiency.