Logical self-reference provides the 'independent network' capable of "learn[ing] to associate specific input patterns with specific patterns of activity of the first network's hidden units."

It is interesting to consider under which conditions a representation will remain unconscious based on combining these two principles (Cleeremans, 2008). There are at least four possibilities. First, knowledge that is embedded in the connection weights within and between processing modules can never be directly available to conscious awareness and control. This is simply a consequence of the fact that consciousness, by assumption, necessarily involves representations (patterns of activation over processing units). The knowledge embedded in connection weights will, however, shape the representations that depend on it, and its effects will therefore detectable – but only indirectly, and only to the extent that these effects are sufficiently marked in the corresponding representations. This is equivalent to Dehaene and Changeux’s (2004) principle of “active firing.”

In closing, there is one dimension that I feel is sorely missing from contemporary discussion of consciousness: Emotion (but see, e.g., Damasio, 1999, 2010; LeDoux, 2002; Tsuchiya and Adolphs, 2007). Emotion is crucial to learning, for there is no sense in which an agent would learn about anything if the learning failed to do something to it. Conscious experience not only requires an experiencer who has learned about the geography of its own representations, but it also requires experiencers who care about their experiences.