From my perspective, there are two basic ideas underlying the overall research strategy entertained by Cruse and Schilling. The first is that in order to understand a system and its properties, it has to be reinvented or reconstructed by the researcher. The second is that mental phenomena may arise as emergent properties via the interaction of low-level components of a system. I’d like to first provide an outline of these basic ideas and the underlying strategy as I understand them. In the next section, I will critically evaluate what types of questions the approach is best suited to answer, and what kind of problems it will likely face.
The first of the two main ideas is central to the research area of bio-robotics. If we are able to create an artificial system that exhibits the phenomena of interest, we should be a great deal closer to understanding how these phenomena come about in nature. In order for this approach to lead to valid conclusions, however, the process of reconstruction has to do justice to the systems we are seeking to understand. In the present context we are concerned, above all, with humans and other animals. This means that the way the artificial system achieves the desired results has to be biologically plausible, i.e., it has to be reasonable to assume that the capacities of the organism that we are trying to understand are really based on similar mechanisms. In this vein, Cruse & Schilling (this collection) are realising the basic reactive modules of their system in form of artificial neural networks that were inspired by biological research on, for instance, stick insects (Walknet) and desert ants (Navinet).
The second of the basic ideas derives its plausibility from an evolutionary perspective on psychological faculties. Emotion, attention, the ability to plan future actions, and any other “higher-level” capacities, including consciousness, did not arise suddenly from one generation to the next and independently of pre-existing, more fundamental abilities, such as the ability to control one’s own body and respond adaptively to environmental stimuli. Rather these latter abilities and the interactions between the mechanisms responsible for them might well have been crucial for mental properties to evolve. From this perspective, the idea of reconstructing the evolutionary process by starting with basic reactive structures and examining how through the interaction of these structures unexpected properties might emerge seems very promising. Since humans also gradually evolved from simpler organisms, it is natural to assume that the same dependence between reactive structures and “higher-level” phenomena is present in our case as well. The investigation of this dependence might thus provide new insights into the mechanisms underlying human psychology.
But what does it mean exactly to say that a property emerges from basic structures? What is an emergent property? The philosophical controversies surrounding the concept of emergence date back over a hundred years and although usage of the term has become increasingly popular in recent years, among both philosophers and scientist, it can hardly be said to have one universal definition. Rather, there are numerous and varied interpretations, a fact which inevitably leads to confusion and misunderstanding (for a good overview see O’Connor & Wong 2012). It is thus vital to identify precisely what is meant by emergence in any particular case. Notwithstanding this inherent ambiguity, there seems, however, to be a shared idea behind much talk of emergent properties: this is the idea that as systems become increasingly complex they tend to exhibit certain higher-level properties, which are novel or unexpected given their simpler, lower-level, components.
Depending on how this claim is interpreted it can have more or less serious implications regarding the fundamental structure of nature, as well as the structure of science. In order to obtain a particularly strong and at the same time highly influential reading, it must be understood in a two-fold sense. First, as meaning that these properties cannot even in principle be predicted or explained on the basis of the lower-level properties of the system and, second, as indicating that such properties are associated with genuinely new causal powers, i.e., they make a real difference to the run of events and are not mere epiphenomena (for discussion see Kim 1999, 2006).[1] This kind of emergence could be called strong emergence.[2] Central to this conception is that emergent properties causally influence the simpler entities from whose organisation they emerge. This sort of causal influence is called “downward causation”, as emergent properties are conceived as higher-level properties arising from certain lower-level properties and relations. Typically, it is assumed that what we find at the lowest level of this hierarchy are the properties and relations of fundamental physical particles. Given this assumption, the existence of emergent properties would entail that a complete description of the fundamental physical organisation of a system might still leave something out. The system might still have some properties that could not be predicted on the basis of such a description and could not be explained in terms of the organisation of its basic physical constituents. Moreover, because emergent properties are causally efficacious, knowledge of the basic physical components of a system and their behaviour may not be sufficient to predict the future evolution of the system. These considerations seem to lead to the conclusion that the meta-scientific thesis, according to which all phenomena can ultimately be explained by the fundamental laws of physics, would turn out to be false. If certain properties belonging to the domains of psychology, biology, or chemistry were emergent properties, these could not even in principle be captured by basic physics alone. All sciences dealing with genuinely emergent properties would remain completely autonomous, positing their own independent laws and explanations. Furthermore, since emergent properties have the ability to causally influence lower-level entities, the fundamental laws of physics would not even suffice to explain processes taking place at the physical level (see also Chalmers 2006).
These are substantial conclusions that could be met with some scepticism. They are also one of the reasons for the fierce controversy surrounding the concept of emergence. Furthermore, the condition that emergent properties are themselves causally efficacious and the general idea of “downward causation” leads to problems in and of itself. This is because there has to be a systematic relationship between emergent and lower-level properties, even though they are conceived as being distinct from another. Often this is expressed by saying that emergent properties are completely determined by lower-level properties and require them for their existence. In other words, if all lower-level properties of a system are fixed, its emergent properties are also fixed; and without any appropriate lower-level properties, a system cannot have emergent properties. If this weren’t the case, it would be unclear in what sense emergent properties emerge from lower-level ones (Kim 2006). If their relationship were completely coincidental, this would surely be an inappropriate description.
Based on this requirement, Kim (1999, 2006) has put forth an influential argument that the idea of “downward causation” is untenable. In summary, Kim’s basic argument is this: suppose an emergent property (let’s say a feeling of thirst) causes a lower-level property (e.g., a certain activation pattern N in the brain). If feeling thirsty is an emergent property, there have to be appropriate lower-level properties from which it emerges. Let’s call these the “emergence base” of feeling thirsty. Now, that feeling thirsty causes N means that there is a natural law that occurrences of feeling thirsty are always followed by occurrences of N (feeling thirsty is nomologically sufficient for N). But since occurrences of feeling thirsty are always accompanied by occurrences of its emergence base, it must also be true that occurrences of its emergence base are followed by occurrences of N. Therefore, if feeling thirsty causes N, its emergence base also causes N. But this makes feeling thirsty completely redundant as a cause of N. Its emergence base is completely sufficient to explain Ns occurrence, leaving the feeling of thirst as a mere epiphenomenon. Since this example can easily be generalised, one can conclude that there are no genuine cases of downward causation and hence no genuine emergent properties of the type presently under consideration.
In summary, it can be stated that emergence is a highly controversial concept—not only because of its inherent ambiguity, but also on account of certain varieties of emergentism that have substantial metaphysical and meta-scientific implications as well as a commitment to the problematic idea of downward causation. The crucial questions remaining now are whether Cruse & Schilling (this collection) provide a clear interpretation of the concept of emergence and whether it provokes the kind of controversy and criticism outlined above. What kind of emergence is involved in their claim that mental states might be construed as emergent properties? In fact, they provide two slightly different characterisations. According to the first, an emergent property is to be understood as a property of a whole system that cannot, at first sight, be traced back to the interactions of the systems components. Alternatively, one might say that we cannot, at first sight, predict the emergent properties of a complex system based on our knowledge of its parts and their interaction. Thus, we might be genuinely surprised that the system in question exhibits such properties. Emergence in this sense is sometimes called weak emergence (Chalmers 2006). If this is all that it means for a system to have emergent properties, few would raise serious objections. This sort of emergence is just a consequence of our limited knowledge and cognitive capacities and is relative to the judging subject: what might not be immediately predictable for one person might be just so for another. Emergentism, in this sense, has no far-reaching metaphysical or meta-scientific implications and is not committed to any sort of “downward causation”.
Cruse & Schilling (this collection) provide a second, and equally unproblematic, definition of emergence that is specifically tailored for application in the context of robotics. According to that definition, a property of an artificially constructed system is emergent if it was not explicitly implemented by its designers. We might call this implementational emergence. This appears to be relatively independent of the sort of “weak” emergence I’ve just described. Even a property not explicitly implemented might be predictable without too much effort, whereas a property deliberately implemented might not be predictable, at least by persons lacking experience or competence. I think that most of the emergent properties Cruse & Schilling (this collection) attribute to their artificial system, reaCog, match both characterisations: they were neither explicitly implemented nor would we immediately expect or predict that reaCog would exhibit them. At the same time, the properties in question are highly interesting and are not simply insignificant side effects. This is important since, according to the definitions provided by Cruse & Schilling, the claim that an artificial system exhibits emergent properties is, in and of itself, not particularly notable. But this depends entirely on what the emergent properties in question precisely are. The finding that reaCog exhibits, in this way, aspects of psychological characteristics, such as emotion or attention and the ability to perform non-trivial body movements, are most certainly of considerable scientific significance. In conclusion, we may say that although the kind of emergentism advocated by Cruse & Schilling does not have the same far-reaching implications as the particularly demanding conception outlined above, it is nonetheless useful and philosophically interesting. This is because it functions as the basis of an intriguing approach to the study of psychological properties, which I shall now endeavour to describe.
Combining the idea of emergence with the idea, outlined above, that in order to understand a system and its properties, it has to be reinvented or reconstructed, we arrive at a fascinating research strategy. The first step consists in observing the behaviour of animals that, although lacking many of the sophisticated abilities with which humans are endowed, are nonetheless capable of flexibly controlling their bodies in order to cope with an unpredictable environment (such as stick insects, desert ants, and honey bees). Based on these observations one then develops a neural network model (e.g., Walknet or Navinet) designed to produce the behaviour observed in the first step. Next, this model is realised in an artificial system (either virtual or robotic) in order to examine to what extent the behaviour produced by the model matches the behaviour of the biological organism on which it is based. If it resembles it to a great extent, this can be taken as prima facie evidence that the mechanisms underlying the behaviour are the same for the animal and the robot. Different modules that are constructed in this way are then integrated into a holistic system. Further modules might be added step-by-step (e.g., Body Model, Attention-Controller, Word-Nets). The result is a complex system (in the present case “reaCog”) the behaviour and properties of which cannot be easily predicted even by its very own designers. The last, and most important step consists of examining whether the system shows characteristics that were not explicitly implemented but instead arise from the dynamic interactions of the system’s components. The most intriguing question in this context is, of course, whether the final system shows aspects of those phenomena that are constitutive of having a mind.
Although this is only a rough sketch of the methodology entertained by Cruse & Schilling (this collection), I hope I have captured the essential points sufficiently to proceed with an evaluation of its scope and the possible problems it might face. What kind of questions is the bottom-up approach best suited to answer? Which phenomena or processes can be addressed by research based on this approach? What considerations have to be taken into account in order for the presented research strategy to be successful? Are there any general constraints bio-robotic bottom-up explanations have to meet? As we shall see, the answers to these last two question are directly connected to two characteristics of the research strategy outlined in the previous paragraph: first, that it involves, at two points, a comparison of the behaviour of significantly different systems and, second, that it is specifically designed to discover emergent properties.