To simulate interference during a delay we present a number of objects to the network'

1
A COMPUTATIONAL MODEL OF RECOGNITION MEMORY IN
PERIRHINAL CORTEX A CRITICAL ROLE FOR COMPLEX
CONJUNCTIVE REPRESENTATIONS
R.A. Cowell 1, S.J. Bartko2, T.J. Bussey2 L.M.
Saksida2, 1. Dept. of Experimental Psychology,
University of Oxford, U.K., and 2. Dept. of
Experimental Psychology, University of Cambridge,
U.K.

• On test, each network was presented with the
  sample stimulus and a novel stimulus in order to
  assess their relative familiarities, as indexed
  by the selectivity of the peak of activation on
  the grid.
• As in rat studies using SOR, we measured relative
  preference for the novel stimulus, indexed by a
  difference score, d (f n)/(f n), where f is
  the selectivity of the familiar stimulus and n is
  the selectivity of the novel stimulus.
• Results

Introduction
Results

• The canonical evidence for a memory impairment
  following brain damage is a delay dependent
  deficit in performance, with good performance at
  short delays.
• Recognition memory tasks such as delayed
  non-matching to sample (DNMS) and spontaneous
  object recognition (SOR) have been used to
  demonstrate a delay dependent deficit following
  perirhinal cortex (PRh) lesions in monkeys (e.g.
  Meunier et al., 1993) and in rats (e.g. Winters
  et al., 2004), respectively.
• In addition, a perceptual role has been proposed
  for PRh (Eacott et al., 1994 Bussey, Saksida
  Murray, 2002) Can we link the findings of
  perceptual and mnemonic contributions of PRh?
• An explanation for the effect of PRh lesions on
  visual discrimination has been proposed in terms
  of complex conjunctive representations in PRh
  that resolve ambiguity between objects with
  common features. Here, in considering the
  mechanism for a delay dependent PRh lesion
  deficit, we apply this explanation to recognition
  memory.
• We have built a connectionist model that can
  account for a delay-dependent impairment in
  memory performance following PRh lesions, and the
  effect of PRh lesions on perceptually difficult
  object recognition memory tasks. The two
  demonstrations employ a common mechanism.

Intact
Lesioned
A two-way ANOVA with Group as between subjects
factor and Perceptual Difficulty as within
subjects factor revealed a significant effect of
Group (F1,10 954.5, plt.0001), a significant
effect of Perceptual Difficulty (F3,30 3761,
plt.0001)and a significant GroupPerceptual
Difficulty interaction (F3,30 274.8, plt.0001).
Connectionist Model
Architecture

• The model comprises two self-organising,
  topographic Kohonen grids PRh and a more caudal
  region of ventral visual stream (VVS).
• It is assumed that representations of visual
  objects increase in complexity with progression
  rostrally down the VVS towards PRh.
• A complex visual object with four features is
  represented by 8 input units, i.e. 2 input units
  per feature. This object is represented as a
  whole in PRh and as four separate features on the
  Caudal VVS layer.

Mechanism
Mechanism

• To simulate interference during a delay we
  present a number of objects to the network.
• We assume that these objects are composed of
  simple features, and that many objects share
  features in common, so that the same features are
  seen repeatedly (and encoded) by the Caudal VVS
  layer during interference.
• The representations of all of these interfering
  features are sharpened.
• However, the probability that a particular
  conjunction of features will occur more than once
  is much lower. The same object is very rarely
  seen more than once by the PRh layer during
  interference, thus representations of interfering
  objects are not sharpened.
• When a novel object, composed of the commonly
  occurring features, is presented to Caudal VVS
  all of its features look familiar and
  discrimination of the familiar object from a
  novel object becomes difficult.
• When a novel object is presented to the PRh layer
  the representation of the familiar object remains
  the only sharp representation.

PRh Layer
Encoding Mechanism
Preliminary Behavioural Data
Two rat groups Sham (sham lesions) and PRh
(perirhinal postrhinal lesions) were tested on
SOR using LEGO objects with two different levels
of perceptual difficulty.

• As a stimulus is presented to the PRh and Caudal
  VVS layers repeatedly during encoding its
  representation becomes sharply tuned.
• In a simulation of SOR the network chooses
  between the familiar and novel objects according
  to the relative sharpness of their
  representations.

A two-way ANOVA with Group and Perceptual
Difficulty as factors revealed a main effect of
Group (F1,22 26.17, plt.0001), a main effect of
Perceptual Difficulty(F1,2235.95, plt.0001) and a
significant GroupPerceptual Difficulty
interaction(F1,2215.88, p .001).

• 2. Object Recognition Memory with Increasing
  Perceptual Difficulty (e.g. Eacott et al., 1994)
• Method
• Intact and Lesioned networks were trained on a
  sample stimulus.
• No interfering stimuli were presented, to
  simulate a zero delay.
• Four Perceptual Difficulty conditions, see
  table below.

Simulations
Summary and Conclusions

• Object Recognition Memory over a Delay (e.g.
  Meunier et al., 1993).
• Assumption the ability of the network to
  distinguish between novel and familiar
  representations deteriorates over a delay owing
  to interference from inputs that intervene
  between encoding and retrieval.
• Method
• Two groups of networks Intact (PRh Caudal
  VVS) and Lesioned (Caudal VVS only) were
  trained on the sample stimulus.
• Four Delay conditions, described by the number
  of interfering stimuli seen during the delay
  0,300,600 or 900.

• The proposed model can account for a
  delay-dependent deficit in object recognition
  memory following PRh lesions.
• In addition, the same mechanism can account for
  the critical role of PRh under conditions of
  perceptual difficulty.
• In each case the critical role of PRh arises from
  its complex, conjunctive representations. Object
  representations in caudal VVS are feature-based
  and provide insufficient information for
  discriminating between stimuli when task demands
  are increased either by introducing a delay or by
  introducing perceptual difficulty.