Hardware architecture for concept learning

One of the hardware architectures used to learn the concepts/vocabulary in my blogs of 1 Oct. and 5 Nov. 2015:

Concept of self

After my AI Asa H developed the 4 levels of concepts as described in my blogs of 1 Oct. and 5 Nov. 2015 I presented the robot with additional activities and a 5th level develops which includes a self concept similar (but not identical) to the one reported in my blog of 4 March 2015.

The AI curriculum again

I have stressed the importance of curriculum when training an AI. The order in which experiences are presented influences the concepts that the AI forms. ("The order that material...is presented...can strongly influence what is learned...sometimes even whether the material is learned at all." pg. ix, In Order to Learn, Ritter, et al, Eds., Oxford, 2007) Cliff Goddard has studied the order in which Wierzbicka's universal semantic concepts are learned by human children (in Applied cognitive linguistics 1, Putz, Niemeier, and Dirven, Eds, Mouton de Gruyter, 2001). I thought this might be relevant for my work on AI curricula (see my blog of 18 July 2014 for example). Some of his results make sense to me.  The concept of "seeing" would be learned early and the concept of "dead/die" would be learned later (at age 3-4 years Goddard says). But some of his results I don't understand.  Why would "hear" be learned so late, for example (at 2.5-3 years of age).  Is it just that we need to be distinguishing between when a concept is acquired and when a word is used to name that concept?