Aaron Gutknecht further proposes a well-chosen list of issues that should be taken into account when following a bottom-up approach as proposed here, namely “adequate matching criteria”, “biological plausibility” and “transparency”.
Concerning the first issue, “adequate matching criteria”, Aaron Gutknecht addresses a possibly critical point. In section 8 (Emotions), we characterize happiness by the property that risky decisions are made more probable. We admit that our example is formulated in a sketchy way, only addressing one basic aspect for illustration. There are, however, more deeply founded examples that have been briefly referred to in the main text and will be explained in more detail here. Two recent studies, one in crayfish, the other in the fruitfly, provide strong hints that emotion-like states can be found in simple organisms as arthropods or, more specifically in the latter-case, insects. In crayfish, Fossat et al. (2014) have convincingly shown that context-independent, anxiety-like behavior can be induced by experimentally applied stress or by application of serotonin. Both methods lead the animals to avoid illuminated sections of their environment which they are normally interested to explore. Anxiety is related to fear but considered a secondary emotion that occurs after the stressing signal has disappeared. Thus, the probability of selecting specific behaviors, in this case exploration of illuminated places, is decreased. This avoidance behavior could be abolished after application of drugs that are known to have anxiolytic effects in mammals. Applied to reaCog, these results could be interpreted in the following way. Emotion-like states would not only influence the global WTA net, but also thresholds of local, lower level WTA networks that are responsible for switching between different procedures.
Another interesting case has been reported by Yang et al. (2014) in Drosophila. These animals learnt that various behaviours selected in trying to avoid a problem, in this case escape from a heated ground, were not successful. As a consequence, they ended up in a state of passivity. This result has been discussed as an example of “learned helplessness”, which is considered an animal model of depression. In our framework, this could simply be realized by freezing activity in the Spreading Activation Layer network that provides input to the WTA net (section 4).
Concerning the second issue, “biological plausibility”, we fully support Gutknecht’s perspective and have only a minor aside. Application of non-spiking neurons is not necessarily biologically implausible. Rather, non-spiking neurons do exist in invertebrate and in vertebrate brains. They play important functional roles, but are generally less well-known, mainly because investigating them involves methodological problems that are more difficult than those of spiking neurons. The third issue, “transparency”, addresses the view that the bottom-up strategy may eventually exhaust its potential when the complexity of the system is further increased. Although we agree with Gutknecht here, we would like to add that the bottom-up approach still bears the advantage that, as the details of such a system are known, its properties can be thoroughly analyzed by physical and/or mathematical methods. This ability, of course, does not guarantee that one will find answers in such a hypothetical case, but there are various methods available to address such questions. Further, we believe that the problem of lacking transparency may not happen to occur too often. This belief is supported by the observation that already our simple system, reaCog, appears to be able to reach integration levels characterized by terms such as intention, volition and consciousness.