This section offers a theoretical perspective on the neurocognition of the representation of action. Applied to rats, it is not to be taken as a theory rival to existing psychological accounts of animal learning, but rather as an account concerning the cognitive representation involved and the cognitive implementation of conditioning. The most prominent feature of the “Cascade” theory of cognitive representation is a multilevel approach to categorization. It applies, it appears, to humans and animals likewise.Footnote 1
5.1 Goldman’s Multilevel Theory of Human Action
5.1.1 Goldman’s Notion of Level-Generation and the Notion of Cascade
When humans categorize and conceptualize an action, they usually do it in more than one way at the same time. The philosopher Alvin Goldman developed a theory of human action that is based on this principal observation (Goldman, 1970). If I open a door, this is a physical act of interaction with an object that changes its state. Opening a particular door can be achieved by a variety of bodily actions. If it is a hinged door, I can push the door at its handle or somewhere else with my hand, I can push it with my foot, I can lean against it with my shoulder or my back; depending on my position and the construction of the door, I may have to pull at the door. For sliding doors or automatic doors, other types of action are required. Thus, ‘opening a door’ refers to at least two levels of action: (1) the basic physical action one applies to the door, and (2) the more abstract functional level of causing the door to open. The acts at the physical and at the functional level do not concern the same properties of the door. The physical act changes the spatial position of the door leaf or leaves. The higher-level act concerns states of the door that are related to its functioning as an object that is used to obstruct or enable access to a space behind it.
The lower-level action is necessary for achieving the higher-level action. This achievement is not automatic but requires certain circumstances; for example, the mechanical door must not be locked, the automatic door must be in function. Goldman (1970) speaks of “level-generation” if actions are related in this way: under certain circumstances, the lower-level action “generates” the higher-level action, the lower-level action is a method of doing the higher-level thing; by pushing the door or pulling at it, one opens it. While in this case the level-generating relation is based on causation, there are also other mechanisms such as conventional level-generation; for example, if I nod my head, this may conventionally generate an approval or permission because nodding one’s head is a conventionally established method of approving or permitting.
Crucially, if an action A generates a higher-level action B, A and B are actions by the same agent and at the same time, done in one. It is very important to note that level-generation does not relate an action to an event it causes. If I open the door for someone and let them pass, I first open the door and then the other will pass through the door a moment later. Level-generation does not obtain between these consecutive actions by two different agents. Rather it obtains between the action of opening the door and the action of opening a passage for the other. These two actions are actions by the same agent and they occur at strictly the same time. It is this feature of Goldman’s theory of action that makes it a theory of multilevel categorization.
According to Goldman, a basic action may level-generate more than just one higher-level action; it can generate a complex multilevel structure of actions with many steps that build on each other; the structure can also branch into different lines of generation. For example, by pushing a door and opening it, one may at the same time open a passage in an aisle as well as cause an air draft; opening the passage may in turn generate doing a favor; causing a draft may further generate making a window slam. We will give complex examples below. Goldman uses the term “act-trees” for structures created by level-generation; we prefer to call them “cascades” as there are good reasons to transfer the notion to other things than actionFootnote 2. Crucially, the actions that form a cascade are actions of different type. For example, leaning against the door and opening the door are not actions of the same type. A door can be opened by other methods, and leaning against the door can have other effects than opening it; for example, it may as well be an act of closing the door, or of keeping the door closed if somebody is pushing against it from the other side.
5.1.2 Goldman’s Level-Theory as a Psychological Theory of Categorization
It is convenient to use the term ‘doing’ for that to which a cascade description applies: there is one doing, for example with the door, but this one doing can be categorized in many different ways as constituting as many different types, or categories, of action as the cascade provides. In the discussion of his theory of action with other philosophers, Goldman emphasizes that the distinction of types involved in a cascade of action is a psychological distinction, not a distinction of things out there in the world. The cascade agent produces one doing, but it is categorized simultaneously at different levels in a hierarchy of level-generation (Goldman, 1979). A cascade forms in our minds, in our view and cognitive modeling of what is going on or what we are doing ourselves. What a person does in a concrete situation, to us, is, in our reality, all these acts in the cascade at the same time. The particular doing in our door example of level-generation may belong at the same time to the action categories ‘push against the door’, ‘open the door for Adam’, ‘do Adam a favor’, and maybe others. It is important to realize that the different categories we may apply to the one underlying doing are not just a bunch of categories that are somehow associated. Rather they are organized in a tree structure of dependence. The higher actions depend for their coming about on the lower actions that “generate” them. And all higher-level actions depend on necessary circumstances to come about.
As the door example illustrates, the formation of cascades takes place even with as simple actions as opening a door. We may well assume that humans categorize almost any willful action by a human as a cascade of action rather than just as the basic physical doing. We will inevitably try to interpret the actions of others in terms of the intentions they pursue by doing what they do; if they act on an artefact in a normal way, for example on a door, we will assume that the action is related to the usual function of the object. Thus, categorizing an action as ‘opening the door’ would provide a causal explanation of the observable physical act.
5.1.3 Social Action and Interaction
One observation relevant in our context is the fact that social action necessarily constitutes higher-level action. Searle (1995) developed a theory of social reality that distinguishes between a physical level and a social level of action, persons, and objects. A certain movement with the head is an approval if and only if it counts as such; a human is the president of Canada if and only if they count as such, and a piece of paper is money if and only if it counts as money. The things that count as something in these examples are physical entities and what they count as are social, entities that is, entities in our social reality. Notably, in all these cases, the things considered are necessarily both at the same time: the physical entity and the social-reality entity. For the part of Searle’s theory concerning acts, the relation between things at the physical and at the social level is captured by Goldman’s more general notion of level-generation.
As a consequence of the principal higher-level character of social action, social behavior always ‘parasitizes’ on more basic physical behavior.Footnote 3 For example, one may turn up the corners of one’s mouth and expose the front teeth and thereby level-generate a smile which, if directed at someone, may under circumstances constitute a social signal which constitutes a display of affection, or something else. Up from the level ‘smiling at someone’, the cascade reaches a social level. If we go back to the example in the introduction, we get an even more complex structure. Using an upward arrow ↥ for level-generation, we can represent the cascade bottom-up as in Fig. 7.
As the example illustrates, own action may cascade to ultimately giving oneself a pleasure (or any other kind of emotional experience) by doing what one does. Obviously, this too cannot be done without the support of some physical action. We may keep in mind two general points about cascades: (i) physical action may cascade to social action, and (ii) action may cascade to obtaining an emotional experience, where emotional-level action may or may not come about by means of social-level action like in our fictitious example.
There is another aspect to the door-opening example. Social reality is constructed interactively (see Clark (1996) on a multilevel interactional account of verbal communication). If A keeps a door open for B, meaning to do B a favor and cascading the conceptualization of their own act correspondingly, then the thanks A receives from B will confirm that A and B share the social construal of their interaction: B would not have thanked A if B had not construed A’s act as involving the level-generation of doing B a favor. An analogous consideration applies to the next level above the favor: the level-generation of obliging B by doing B a favor. Acknowledgment and confirmation of this additional, emotional level, is executed by B sending a smile to A. Given that receiving a smile is felt as something pleasant, the level-generation of ‘B please A’ by ‘B smile at A’ is part of the joint construal of B’s reaction.
5.2 Cascades and Learning
Goldman’s theory was constructed for the categorization of individual action tokens. It can, however, also be applied to the consistent multilevel categorization of recurring types of action. For example, if we experience that the light goes on when we flip a certain switch, and if we repeat the action and achieve the same effect, we will learn a cascade: that flipping this switch goes with switching that light on. We acquire a piece of procedural knowledge by memorizing a two-level action cascade concept composed of the two single-level action concepts ‘flip this switch’ and ‘turn on that light’. Our environment being as it is, we will easily generalize this cascade to other switches and other lights, and so on. Thus, action learning is cascade learning, at least for all but merely physical basic action like turning one’s head or lifting ones hand. We learn that doing one thing also means doing the higher-level thing, and the level above that, and so on. An action and the higher level achieved with it are conflated into one concept. Cascade learning may also include that, by an action, we trigger approval or disapproval, cause pleasure or pain, a particular taste or other bodily sensations. If we assume that cascading plays a role in concept formation, we may conclude that action concepts are formed that link basic actions and the recurring achievement of certain causal effects into one multilevel concept.
It is important to note that even for humans, kids or adults, learning of action cascades does not necessarily involve reflection. It just requires that the learning subject register that the lower-level action goes with the higher-level action. In particular, the learning of cascade levels that are causally linked does not require any causal understanding. We learn that pressing the red button of the TV remote control means turning the TV on or off, but we may well die without ever having understood what we actually do at the technical level by pressing this button. This level of understanding is not relevant for learning how to succeed in dealing with TVs and remote controls. To know how to deal with a remote control is essentially ‘knowledge how’Footnote 4, and the mechanism by which we acquire this knowledge is learning by doing.
Cascade learning does not only concern practical abilities. A child may cascade-learn that a certain kind of behavior always upsets her mother; the child will register this and adjust her behavior accordingly, but may possibly never understand why her mother reacted that way. We learn in countless regards that our actions are accompanied by higher cascade levels of particular qualities. Cascade learning will result in a “practical” implicit understanding of the environment, in the sense that we learn which intended or unintended higher-level kinds of action are generated by certain other kinds of action. We learn things like “if I do x, I give myself experience y”. Given that we are able to undertake certain action or refrain from it, this kind of understanding our environment will enable us to adapt to it.
Cascade knowledge need not be accessible to consciousness: we may have it without being aware of it and without being able to describe it. For example, pronouncing a word in a way that enables others to recognize it phonetically means to enact a cascade of production based on intentional action of our articulatory organs to produce articulated sounds, thereby producing speech sounds, thereby producing certain speech phonemes, and with them an established sound form of a word in a particular language. All this is stored in the language production repertoire—a normal language user is not aware of the levels of actions involved and they would not be able to describe what they do at which levels. All they need to be able to do is to aim at doing something particular at a pretty high level, something with the result of making audible a particular sound pattern.
5.3 Applying Cascade Theory to Rat Behavior in the Experiments Reported
We proceed to propose that the cascade model of action categorization and action learning applies to rats as well. First of all, it appears that there are certain types of rat action that are relevant to the actors at levels beyond the mere physical doing. Among these are levels that constitute social action. For example, if young rats do rough-and-tumble play, they recognize that this is not hostile fight: crucially, the fight is ‘friendly’ to both of them. In some way or other, they succeed in letting the other “know” that their own behavior is not hostile, and they succeed in categorizing the other’s behavior in the same way. Both rats engaged in a rough-and-tumble play possess two categorizations of representing physical fight or else fight-like action. At a lower level they categorize the physical action, at a higher, social, level they categorize it as hostile fighting or as play. At the lower level they “know” bodily methods of fighting, for instance, pushing or biting, and they are able to modulate these methods as to cascade either to a real fight or to rough-and-tumble play. There can be no doubt that a rough-and-tumble play to the rats is both, bodily interaction and a social interaction that is different from hostile fight. What they do has a function to them at both levels, as some sort of bodily learning and some sort of social learning.
When we say that this “is to the rats” a particular type of action, we do not imply consciousness on the part of the rat. The cascade view does not commit us to the assumption that rats have consciousness (at least not in the same sense as humans); it only commits us to assume that the rats’ brains categorize the rats’ doing in these ways and that, in this sense, the rats register what is going on at both levels. As rats are able to recognize and repeat types of action, for example under experimental conditions, they must have cognitive representations of types of action. Crucially, they register the character of what is going on not only for their own part, but also for the part of their interaction partner.
Among the actions that have multilevel character to rats are the USVs (ultrasonic vocalizations) mentioned above. The fact that these vocalizations trigger brain reactions associated with emotion, shows that these are not just plain sound productions (like, for example, the production of the sound they produce when they scratch their ear or shuffle around); these special sound productions are ‘received’ at an acoustic and an emotional level. We do not know if the rat, when hearing a 50-kHz USV, hears this as a display of comfort or pleasure. If this should be the case, the rat might have a two-level representation of the act by their conspecific. All we seem to be entitled to assume at present is that 50-kHz USVs must have a pleasant ‘ring’ to the perceiver. But this is sufficient for the assumption of a two-level neural cascade representation of 50-kHz USVs issued by other rats, whence these USV’s carry emotional significance.
In the experiments, the rats learn. They acquire behavior. The experiments are designed in the way that the behavior acquired leads to getting themselves a reward. We can apply the cascade model to the learning process, if we assume that learning a particular behavior consists in acquiring a multilevel action cascade. The general structure of reward-inforced learning would be the acquisition of a cascade that amounts to: ‘do x’ ↥ ‘get a reward’; here ‘get’ is to be taken to mean active acquisition, not just passive reception, because the latter would not be an action by the animal.
Assume that the actor rat learns that it will receive pellets upon entering compartment c1. That learnt, the rat will repeat the action if it likes to get pellets. This behavior can be interpreted as involving the acquisition of an action cascade of three levels:
One might speculate that it is the rewarding course of events that supports not just the behavior as such, but primarily the formation of the cascade described; if the animal forms and then memorizes the cascade, this results in a mental condition that enables the animal to repeat these rewarding experiences at will by taking the action at the bottom of the cascade.
In the prosocial choice task experiments described in Hernandez-Lallement et al. (2015), some rats seem to have learnt just Cascade 1. The prosocial rats, however, developed a behavior that involves a more complex cascade structure with a second branch on the first node (Fig. 8, blue branch).
They register that the partner rat gets pellets, too, and their brain ascribes it to themselves as a generated higher-level action. As for the third step of the cascade, we know that there are 50 kHz USVs when the actor rat and the partner rat simultaneously get their pellets; however, due to the technical equipment used, it was not possible to ascribe the vocalizations to one or the other rat or to both. We are entitled to assume that the actor rat sees and thereby registers that the partner rat gets pellets. We do not know whether this constitutes a pleasant experience to the actor rat. If we could be sure that the partner rat produces a 50 kHz USV, we might assume that the actor rat hears it and experiences this as an emotional reward. We can explain the preference for this condition only if we assume that the prosocial behavior cascades to an additional reward in the left branch of the cascade. The left cascade branch would then level-generate an additional third-level ‘get pleasure’.
In the new experiments described above, with no partner rat present, the USV constitutes an additional two-step branch generated by Level 2 in the first cascade, to be construed as ‘get a 50 kHzUSV’, thereby ‘get pleasure’. We will assume that the two-way reward (transiently) outweighs the one-way reward of the no-USV condition. An explanation as to why the effect of the USV gets weaker in the course of the experiment will not be attempted here.
5.4 Psychological Commitments of the Cascade Approach
Application of the cascade model to rat learning involves certain psychological commitments.
The rat’s brain implements cascade formation.
The rat’s brain creates links between basic types of bodily action and what goes with them, perceptibly to the rat; if the rat brain works in this way, it ascribes the effects of behavior to the behavior itself, connecting, for example, eating certain food to staying hunger. In this way, the animal learns by experience what its behavior “means” to it. When we talk of “meaning” here, we mean it in the basic sense of immediate concomitance, not involving reasoning or convention: if something is of category A and category A cascades to category B, then this instance of A “means”/“constitutes”/simply “is” also an instance of B—to the cognitive subject.
Level-generation presupposes that the animal perceives its own action, and attributes it to itself. This results in a second psychological commitment:
Rats have a (weak) sense of agency. Their brain records their action.
There can be no doubt that rats, by way of proprioception and perception of the environment, sense that they are acting.
In addition, we need to assume that the rat’s brain categorizes what the animal is doing. This amounts to the following commitment:
The rat’s brain forms concepts (representations) of types of own action.
Crucial for the cascade formation is the following assumption:
The rat’s brain assigns credit to the animal itself for what happens concomitant with action of the particular type.
Cascade formation then means that the rat’s brain generates a higher cascade level for the underlying action concept that amounts to making happen what happens after action of the given type.
There are restrictions on this condition. First, we will assume that it holds only for such events following the rat’s action that are significant to the animal’s well-being, and hence of “interest”. Second, there is supposed to be a limit on the time that may lapse between the rat’s action and later events. The rat’s brain will possibly not connect the animal’s doing to things that happen after a long time.
It appears that commitments (ii) to (iv) are uncontroversial; we construe the changes in behavior of rats in experimental settings as learning behavior under the conditions of the experiment. This would be unexplainable if we would not assume that the animals’ brains register the animals’ doings as their own and as of a particular type and if their brains did not credit the animals with what follows their own action as something they can ascribe to this type of action.Footnote 5
The rat’s brain stores in long-term memory the repeated concomitance of certain effects with a type of own action.
This means that the rat’s brain connects this type of action—not only individual single action tokens —with this kind or result.
Of course, the crucial assumption is the first one. The other assumptions are implicit in everyday experimental practice.
5.5 What Can the Cascade Approach Buy Us?
What the cascade theory buys us is twofold. First, it provides a fundamental neurocognitive mechanism for a model of the animal’s learning about its environment. If the animal’s brain builds cascades on the types of physical action the animal is capable of, then the brain integrates the type of action with the achievement of its results into one multilevel concept. The type of action is thereby invested with a particular significance for the animal, for example emotional significance, significance relevant for survival, or the significance of performing a certain type of social action or interaction. Cascade-format action concepts link an action to the achievement of its result as something ascribed to the animal as self-caused—and thereby controlled by own behavior. Cascade learning of effects of their doing invests the animals with the ability to choose ways of action, to seek advantage and avoid disadvantage. Thus, cascade formation for own-action types provides a basic mental mechanism of adapting to the environment, including the animal’s social group, in a learning-by-doing way.
Second, the cascade approach offers an explanation for the way in which an animal is able to acquire a practical understanding of the ways of its environment, as its brain links types of behavior to the triggering of its outcome. If the cognitive system of an animal is equipped with the ability of action cascade formation—i.e. if it records what the animal does to itself if it acts in this way or another—it enables adaptation to the environment without requiring any level of causal understanding, reasoning, or modeling. Thus, the cascade model of learning is a model of learning by doing and what is acquired is plain knowledge-how.
The cascade approach might be successful in modeling multilevel categorization across humans and animals, in particular as part of modeling the acquisition of multilevel action concepts and methods of how to do things, and of what is ‘social reality’ to the cognitive subjects. Another way of looking at cascade theory is to consider it a psychological theory of “meaning”, in the very basic sense that acting at a lower cascade level also “means” to act at the generated higher level. In this sense, cascading provides action with meaning to the cognitive subject.
In the field of cognitive theory and psychology, the theory is at its very beginning. It seems to be able to claim some plausibility (cf. Vallacher & Wegner, 2011). In any event it would be interesting to try to develop methods for testing it experimentally. For example, a cascade approach to learning raises concrete questions concerning structural constraints on cascades to be acquired in terms of the number of levels, of branching complexity, and of memorizability.