Talk:Confabulation theory (computational intelligence)
Review #2
This article appears well written and clearly presents the important aspects of the theory.
It is important to note that Confabulation Theory is motivated by experimental results in neuroscience, but many of the key ideas have not yet been rigorously tested in biological systems. Computer models have confirmed the utility of confabulation as a mode of information processing, however, which makes the theory of interest even if future experiments disagree with postulates made by Confabulation Theory. Biologically-minded readers should note that many of the statements in the article are not claiming to be experimentally-grounded, biologically accurate accounts. They are postulations, and are identified as such.
Unfortunately, the "Some Predictions" section is currently rather weak. It is adequate for the article, but could use further refinement to include more examples of specific observations one might try to make to confirm or reject specific aspects of the theory. For example, the two-stage synfire chain proposal may suggest that for every neuron that represents a pattern, there is a set of relay neurons that reliably transmits the activity of the original neuron. For neurons with easily-identified receptive fields, such as those in the visual or olfactory cortex, one may therefore expect to find relay neurons with similar receptive fields (but possibly different properties). Expanding upon issues like this would be most helpful in a section on biological predictions.
Review #3
This scholarpedia article on "confabulation theory" gives a good overview and introduction on confabulation theory. It is clearly structured and well readable. Perhaps the last section on "confabulation technology" could be improved and extended. For example, one might be further interested more specifically which predictions could be tested by which particular experiments. I am also not sure whether "confabulation technology" is here an adequate chapter heading.
In the section on "confabulation" I was not so clear about the relation between the so-called confabulation operation (the WTA retrieval) and the probabilistic interpretation at the end of the section (i.e., the claim that the "cogency" pr[alpha..delta|eps] is maximized). For me this relation is not obvious because it depends on how the synaptic weights of the "knowledge links" are defined. Perhaps it would help to specify this already in the section on the "knowledge links". For example, my default assumption was that the weight from alpha to eps should be something like pr[eps|alpha]. But then I realized that this will not lead to the claimed cogency maximization. However, it might work with log pr[alpha|eps] or log pr[alpha,eps]. In any case it would help to spend a few words on this issue either in the section on "knowledge links" or in the section on "confabulation". This will also lead to further predictions on synaptic learning which should possibly be included in the section "Some predictions on CT".
There are also other closely related large-scale associative brain theories which should be cited, for example: Knoblauch, A., Markert, H., Palm, G. (2005) An associative cortical model of language understanding and action planning. Proceedings of the International Work-Conference on the Interplay betweeen Natural and Artificial Computation. Lecture Notes in Computer Science, Vol.3562, Springer-Verlag: Heidelberg
Also I did not find the citation (Sejnowski and Destexhe 2000) in the references. Please add.
A technical sidenote
Embedding text in an image, as done in the side pannel is problematic because search engine robots and blind/disabled people who use voice synthesis or tactile displays are unable to get the content. This is considered bad practice on the Web
