Abstract
Objects are not stored as unitary representations: memory representations are hierarchically structured in that objects structure memory, but features within objects are mostly independent (Fougnie et al. 2013). Yet prominent chunk learning theories still treat learned chunks as all-or-none bound representations (e.g., in visual memory, “content-free labels”, Huang & Awh, 2018). Here we suggest that as in the case of objects themselves, chunks that combine objects are also hierarchically structured, with preserved information about their contents in addition to the chunk ‘pointer’. We tested this in two ways: In Experiment 1 and 2, we had participants learn new ‘chunks’ using the color-pair paradigm of Brady et al. (2009), in which colors frequently co-occur next to each other in a working memory paradigm (e.g., orange frequently co-occurs with green). However, unlike previous work where the colors were identical on every trial, the colors in these experiments varied in luminance: so although green frequently occurred with orange, the particular green and orange on each trial were unique. We found that (1) learning significantly improved capacity over the experiment; (2) after learning, knowledge about the 2 colors was correlated, showing some binding; and yet (3) participants did not simply encode abstractions (“irish flag colors”), but had preserved knowledge of trial-specific color information as well. In Experiment 3, we relied on existing chunks. Participants were instructed to remember known words (e.g., “GO”) or unknown letter pairs (e.g., “SP”), with letter fonts varying from trial to trial. The presence of pre-existing word chunks helped not only in knowing the letters but also improved knowledge of the fonts -- which contradicts predictions made by a content-free labels account. Overall we find evidence that memory should be considered hierarchical: Chunks, like objects, serve as cues to access specific information within them, rather than serving as all-or-none abstractions.