Well that’s good evidence that recall is constructed, isn’t it?
If events were stored as abstractions, wouldn’t that make them harder to change?
Current NN’s are notoriously hard to change. The whole thing needs to be retrained from scratch at the addition of every new bit of information. Figuring out how to change them incrementally is another big research conundrum.
For myself, because I see meaning as a constant process, I like to distinguish “recall” from “memory”. The exact words don’t matter. But I like to emphasize that what we can become consciously aware (recall) of might be different to the storage underlying that awareness (memory.)
But, call it recall or memory as you like, if subjects were susceptible to having their recollections gradually altered by new details supplied by the experimenters, that to me says that the recall process is ongoing, and susceptible to both the constant addition of new elements and new organizations of existing elements.
By the way, your mention of single word utterances as the basis of normal speech development, in contrast to unanalysed chunks more typical of autistic development, motivated me to look up some early childhood learning data. Am just reading this interesting discussion:
'Unanalyzed imitative routines are one form of gestalt processing found in normal children’s language. …
The use of memorized, unanalyzed segments has also been studied for normal subjects. Peters (1977) discussed a gestalt style of language use wherein some early language learners memorize particular multiword phrases that are heard often in specific contexts. She suggested that such phrases may actually be perceived as single units and may subsequently be used somewhat appropriately in similar situational contexts, giving the appearance that the linguistic system is of greater complexity than it actually is. Peters suggested that children who demonstrate such gestalt language may initially have difficulty using pauses and other prosodic cues to segment utterances. As a result, they pro-duce whole utterances rather than one or two words.
Peters’ observations are supported by Clark, who indicated that her sons used copied utterances that “were retained in tact for several weeks,” such as “wait for it to cool,” which was said when a hot meal was brought to the table (1974, p. 4), or “Don’t touch that, it’s hot,” which was said as one child pointed to hot tea (1980, p. 10). These utterances were far more grammatically sophisticated than the children’s actual level of linguistic competence; thus Clark referred to the production of such patterns as “performing with-out competence”’
I also recall evidence that children’s production of English past tense (one of the early successes of connectionism) is famous for initially producing “irregular” forms correctly, then over regularizing them, before going back to idiosyncratic irregular forms again. Suggestive to me that initially these are stored as unanalysed chunks, which then get over regularized in a productive abstraction process, before the abstraction process is constrained again. This paper seems to provide a nice discussion of that (and more?):
Children’s Acquisition of the English Past-Tense: Evidence for a Single-Route Account From Novel Verb Production Data
Ryan P. Blything, Ben Ambridge, Elena V.M. Lieven
I also don’t know if it’s worth mentioning again in this context, the analysis I referred to before by Peter Howarth, looking at the normal progression of competency in second language learners, maybe different from children, or maybe related. Howarth found that learners tended to produce what he called “overlaps”, or mixes by analogy between detail sequences. Only later did they start to produce more of what he called “blends”, which were generalizations based on what might be called more abstract generalizations:
I’d be interested to hear your suggested methods.