Pentti Haikonen comments invited by Conscious-Robots.com editor on the recent article about Haikonen’s Architecture for Conscious Machines written by Trung Doan.
I wish to thank Trung Doan for his analysis of my approach towards machine consciousness here. Trung Doan has done a lot of work here as he illustrates the main principles with practical examples, which are worked out along the principles that I present in my book “Robot Brains”.
I would like to add some comments. My realization relies on associative neurons, which form associative memories. An associative memory is a rather old invention, but it has not gained much popularity because of the so-called interference problem, which limits the capacity of the memory. My contribution relates to the interference problem and in my book I describe methods, which allow the interference-free use of the full capacity; the capacity of the associative memory can be the same as the capacity of similar complexity random access memory. However, if only partial capacity is used, then the associative memory also performs the act of classification.
Trung Doan notes correctly that the neural machinery does not operate with numeric values, instead the individual signals represent elementary features of sensed entities and these are the basic meaning of these signals. It is useful to note that in the machinery groups of signals that represent some sensed entity may be used to stand for completely different things; these signal groups act as symbols for these things. This is a necessary prerequisite for e.g. natural language and inner speech.
Trung Doan’s analysis of pain and pleasure in the machine needs some clarification. Pain and pleasure are system reactions. Pain reaction is triggered by signals from damage detectors. This reaction operates via threshold controls in such a way that it will try to interrupt all on-going activity so that other activity might intervene. Eventually some emerging activity may remedy the situation and pain will cease. There are no rules (apart from some hardwired reactions) what the pain should do. Pleasure reaction tries to sustain on-going activity and operates also via threshold control. Again, there are no rules. (Rules are needed, if one simulates the operation of the system by a computer program). The machine will learn on its own to avoid pain-producing activities and to seek pleasure-producing activities.
In humans and animals “good” and “bad” concepts originate initially from taste and smell sensors. A robot may not have such sensors and therefore artificial “good” and “bad” inputs may be used; these would be used for reward and punishment purposes.
Is the Haikonen machine conscious? It may be, because it perceives its environment and own body and it is able to introspect its mental content.
However, all this can be subconscious; after all most of our own mental processes are subconscious. But, the difference between conscious and subconscious is reportability. The Haikonen machine has this reportability.
The various sensory and motor modalities can report their content to other modalities, which then may act according to the received information. The machine can also produce reports to outside observers.
However, the final criteria for consciousness is, as I see it, qualia. Does the machine perceive the world via qualia? The qualia does not have to be similar to ours, but nevertheless the qualia must be there. I have discussed the qualia problem in the December issue of the Journal of Machine Consciousness and those interested in this aspect might find something useful there, maybe. Nevertheless, the Haikonen cognitive architecture is designed so that the assumed prerequisites for qualia are met. This, of course, applies only to the hardware realization. Software realizations may be useful and educative, but they will lack qualia, I am sure.