| |
Abstract:
Abstract: Recent studies of recognition memory suggest that,
while the hippocampus plays a necessary role in supporting
recollection of specific studied details, the surrounding medial
temporal lobe cortices (MTLC) can compute stimulus familiarity
(i.e., the extent to which a stimulus resembles previously studied
stimuli) on their own. We use a biologically realistic neural
network model of MTLC to explore this structure's contribution to
recognition memory. Our MTLC model adheres to the principles of the
Complementary Learning Systems framework set forth by McClelland,
McNaughton, & O'Reilly (1995); according to this framework,
MTLC (and the rest of neocortex) use overlapping representations
and learn slowly, in order to extract reliable statistical
relationships between features in the environment (in contrast to
the hippocampus, which learns rapidly and uses relatively
non-overlapping representations to mitigate the interference that
would otherwise accompany rapid learning). Using our model of MTLC,
we show that -- even though MTLC learns slowly -- small weight
changes associated with prior study have reliable effects on the
sharpness of representations in MTLC: Familiar stimuli strongly
activate a small number of units, whereas unfamiliar stimuli weakly
activate a larger number of units. Finally, we show how our model
of MTLC-mediated familiarity can explain several challenging
recognition findings from normal subjects and subjects with focal
hippocampal damage (e.g., the null recognition list strength effect
reported by Ratcliff, Clark, & Shiffrin, 1990).
|
