Review

Letter perception:
from pixels to pandemonium
Jonathan Grainger, Arnaud Rey and Stéphane Dufau
CNRS and Aix-Marseille University, Laboratoire de Psychologie Cognitive, Université d’Aix-Marseille I, 3 place Victor Hugo, 13331
Marseille, France

In 1959, Oliver Selfridge proposed a model of letter cognitive theory of letter perception. In Pandemonium,
perception, the Pandemonium model, in which the cen- letter identification is achieved by hierarchically organized
tral hypothesis was that letters are identified via their layers of feature and letter detectors. Support for such a
component features. Although a consensus developed hierarchical organization was provided at that time by
around this general approach over the years, key evidence neurophysiological studies of the cat visual cortex [8]. In
in its favor remained lacking. Recent research has started spite of this evidence, an alternative theoretical approach,
to provide important evidence in favor of feature-based template-matching, has also been favored by cognitive
letter perception, describing the nature of the features, scientists. Template-matching models assume that several
and the time-course of processes involved in mapping shape-exemplars of a given letter are stored in memory,
features onto abstract letter identities. There is now hope and that recognition consists in finding the best match
that future ‘pandemonium-like’ models will be able to between a target item and one of these memory traces. In
account for the rich empirical database on letter identifi- this approach, a new template is learned and stored each
cation that has accumulated over the past 50 years, hence time a new target stimulus differs notably from existing
solving one key component of the reading process. templates. Simple versions of template-matching compare
descriptions of the stimulus as a set of pixel intensities
What is the letter ‘a’? with corresponding representations in long-term memory,
The cognitive scientist and philosopher Douglas Hofstadter but the distinction between feature-based and template-
once noted [1] that ‘The central problem of Artificial Intelli- matching approaches can be blurred by incorporating tem-
gence is the question: what is the letter a?’ What Hofstadter plate matchers as feature detectors in Pandemonium [9].
was suggesting in his provocative statement is that un- Nevertheless, the major drawback with template-match-
derstanding the mechanisms underlying invariant recog- ing models, as already noted by Neisser [10], is that the
nition of the arbitrary signs that compose the Roman matching procedure requires prior normalization of the
alphabet (a = A 6¼ b) will be a major step towards under- stimulus (adjusting the stimulus to a prototypical position,
standing the essence of human intelligence. Letters size and orientation), and proposals for such a process in
represent a perfect example of the kind of symbol that most cases lack psychological and neurophysiological
humans thrive on, and letters are sufficiently limited in plausibility. Furthermore, a general consensus has devel-
complexity and number to provide a highly tractable domain oped over the years in favor of feature-based approaches.
of investigation. Letters are also the gateway to reading [2– What is the key evidence for this, and what are the
4], perhaps the most complex skill that humans have to features?
master without specific genetic predisposition. Further-
more, a letter-based strategy for reading in alphabetical
orthographies has probably developed because it is far more Show me the features
economical to solve shape invariance for 26 letters compared The confusion matrix is the traditional method used to
with tens of thousands of words [3], and understanding hunt for features. In a typical experiment used to generate
shape-invariant recognition is a major endeavor of current a confusion matrix, isolated letters are presented in data-
research on visual object perception [5]. In this article we limited conditions (brief exposures and/or low luminance
review exciting new developments in this central topic of and/or masking) and erroneous letter reports are noted.
cognitive science. Recent research provides converging evi- Error rate (e.g. reporting F when E was presented) is
dence in support of the classic account of letter perception hypothesized to reflect visual similarity driven by shared
formulated by Oliver Selfridge 50 years ago (the Pandemo- features. An analysis of the pattern of letter confusions was
nium model), and the hope that a complete account of the therefore expected to reveal the set of features used to
processes involved in recognizing a letter of the alphabet is identify letters. There are >70 published studies on letter
within reach. confusability, and some have formed the basis of concrete
proposals of lists of features for letters of the Roman
Pandemonium in the air alphabet, mainly consisting of lines of different orientation
The starting point of contemporary research in this field is and curvature [11–13].
Selfridge’s [6,7] seminal work, laying the foundations for a One major drawback of standard letter confusion data is
that the method used to degrade stimuli (to generate
Corresponding author: Grainger, J. (jonathan.grainger@univ-provence.fr). confusion errors) influences the nature of the confusions
1364-6613/$ – see front matter ß 2008 Elsevier Ltd. All rights reserved. doi:10.1016/j.tics.2008.06.006 Available online 27 August 2008 381
 Review Trends in Cognitive Sciences Vol.12 No.10

Figure 1. Illustration of inter-letter similarities (Euclidian distances) revealed by a same–different matching experiment [16]. (a) A dendogram plot in which the height of
connecting bars reflects Euclidian distance (smaller = closer). Inter-letter similarities form clusters shown in different colors. (b) A projection of the same Euclidian distances
on a plane [44]. The original distances in high-dimensional Euclidian space are coded by the color of connecting lines. In both panels it can be seen that letters with specific
combinations of simple features group together, such as ‘vertical line’ plus ‘circle’ (b, d, g, p, q). This corresponds to the main similarity class obtained by principal
component analysis. The second similarity class, corresponding to small curvilinear shapes (a, c, e, o, s), is also visible, in addition to combinations of similarity classes
such as ‘vertical line’ plus one small feature (f, t, i, j, l).

and furthermore confounds perceptual confusions with were clearly interpretable as elementary visual features
post-perceptual guessing. In one of the first studies to (Figure 1).
overcome this drawback, Podgorny and Garner [14] used
a same–different matching task and showed that the A new look at features
resulting discrimination-time matrix for letters correlated Two recent articles open up an exciting new perspective for
well with judgments of perceptual similarity (see also Ref. research investigating feature-based letter perception. Pelli
[15] for confusion matrices expressed as saccade latencies). et al. [18] measured contrast thresholds to Roman letters
Furthermore, having response time (RT) as the dependent presented in different fonts, in addition to letters and char-
measure avoids the problem of empty cells and provides a acters from other languages. The authors expected letters to
ratio scale that enables the application of more powerful be identified optimally, like a single feature (i.e. a spatial
metric analyses. More recently, Courrieu, Farioli and frequency channel or filter in a grating experiment). This is
Grainger [16] used a Go–NoGo variant of the same–differ- what would be expected from a template-matching
ent matching task, with participants responding only when approach. To test this, Pelli et al. [18] measured efficiency
the two letters were different. The discrimination times of letter identification (Box 1) under varying viewing con-
were transformed into Euclidean distances by using a ditions and found that efficiency was independent of
‘monotonic embedding’ technique [17]. A principal com- stimulus duration, eccentricity and size, but did vary across
ponent analysis revealed 25 dimensions, many of which different alphabets and fonts. The sub-optimal performance

Box 1. The psychophysics of letter perception

The visual world can be described in terms of variations in spatial frequency of the channel varying as a function of letter size, font
frequency, that is, changes in luminance across space [45]. Letters and alphabet [21,49]. This could be because the presence of noise in
have a broad spatial frequency spectrum, so what part of the the critical-band masking procedure makes it harder to learn to use
spectrum is used to identify letters? Two main techniques have been different spatial frequency channels in the course of an experiment
used to answer this question – measuring identification thresholds for [50] or because filtered stimuli force participants to use information
bandpass filtered letters, and measuring variations in identification that they would otherwise ignore.
thresholds as a function of the bandpass characteristics of a masking This line of research was extended further in the recent work of Fiset
stimulus (critical-band masking, Figure 2 in main text). The key to et al. [20] applying the ‘bubbles’ technique of Gosselin and Schyns to
both approaches involves comparison of human performance with letter identification. The bubbles technique aims to uncover the parts of
that of an ideal observer. Letters are better identified with high-pass the image that are diagnostic of the observer’s performance (i.e. correct
filters than with low-pass filters, but this is because more information letter identification). The technique, therefore, not only manipulates
is available in the high-pass stimuli. Ideal observer analysis enables spatial frequency filtering, but also explores the significance of different
information availability to be equated, hence enabling biases in parts of the image. As illustrated in Figure 2 (In the main text), the image
information utilization to emerge [46]. The measure of human is first band pass filtered at different channel frequencies and a random
performance relative to the ideal observer is called efficiency. set of ‘bubbles’ is extracted from the filtered letters. The sampled
On the one hand, several studies have reported that observers images are summed across frequency channels to generate a
identify octave-band wide filtered letters with almost as much ‘bubbelized’ image presented to observers. Variation in performance
efficiency as unfiltered letters for all but extreme (very high and very is then traced back to information available in the image (coordinate of
low frequency) bandpass filters [46–48]. On the other hand, critical- bubble center) at each frequency band using image classification
band masking experiments indicate that only a single channel is used procedures (least-squared multiple linear regression on spatial co-
to identify broadband (i.e. unfiltered) letters, with the centre ordinates and performance).

382
 Review Trends in Cognitive Sciences Vol.12 No.10

Figure 2. (a) Illustration of critical-band masking. Broadband letters (top row) are combined with bandpassed noise (middle row) to form a masked letter stimulus presented
for identification. Three cycles per letter noise (middle) acts as a better mask than masks formed by either lower (left) or higher (right) spatial frequencies. (b) Illustration of
the ‘bubbles’ technique applied to letters. Letters are first bandpass filtered with a series of five octave-wide spatial frequency channels (top row). Each filtered image is then
randomly sampled (randomly selected x and y coordinates), and each sample smoothed by a Gaussian kernel with standard deviation proportional to the channel
frequency. This generates a set of ‘bubbles’ representing parts of the image (middle row), the number and size of which is determined by the channel frequency of the
filtered letter in the top row. The bubbles are then used as a spatial window that determines what is extracted from the filtered images (top row) to form the parts of the
image shown in the bottom row. The sum of these image parts forms the final bubbelized image of the letter A (right image). Adapted, with permission, from Ref. [20].

of human observers was therefore taken to reflect feature- diagnostic piece of information for letter identification,
based letter identification, in which identification of the with intersections and horizontal lines providing further
whole is affected by the identification of each component significant sources of information for uppercase letters. For
feature. What is it that changes across alphabets and fonts example, the letter W was mainly distinguished from other
that might be driving these changes in efficiency? Pelli et al. letters by the presence of two terminations, one in the
[18] found one particular measure that correlated highly upper left corner and the other in the upper right corner.
with letter identification efficiency. That was perimetric
complexity – the square of the length of inside and outside The time-course of letter perception
perimeter, divided by ink area (for size invariance). In the Standard behavioral measures of letter identification, such
absence of independent evidence concerning the nature of as percentage of correct responses in data-limited con-
the features subtending letter identification, perimetric ditions and RTs to non-degraded stimuli, all represent
complexity provided a measure of visual complexity thought the final result of an accumulation of component processes.
to be proportional to the number of features. However, a complete understanding of letter perception
The second breakthrough has come from recent must incorporate knowledge about how the component
research applying Gosselin and Schyns’ [19] ‘bubbles’ tech- processes develop through time. The masked priming
nique (Box 1 and Figure 2) to explore the nature of the paradigm [22] has proven its utility as a tool for examining
critical features for letter perception. The classification the earliest phases of visual word recognition (see Ref. [3]
images obtained by Fiset et al. [20] for 26 lowercase and for review), and has been usefully applied with letter
26 uppercase Roman letters in Arial font revealed several stimuli to separate out the role of visual factors from
important pieces of evidence. First, on average only 32% of phonological and articulatory factors in letter identifi-
the printed area of uppe and 24% of lowercase letters was cation and letter naming (Box 2). Masked priming can also
used by observers to identify letters, and the greatest be combined with measures of brain activity that provide a
proportion of useful information was apparent in the 2–4 moment-to-moment reflection of on-going target proces-
cycles per letter frequency band, in line with estimates sing. In a recent study this combination was used to
from critical-band masking studies [21]. Second, the investigate the time-course of component processes in
analysis revealed that terminations were by far the most letter identification [23]. This work revealed a cascade of

383
Review Trends in Cognitive Sciences Vol.12 No.10

Box 2. Masked priming with single letters

In the masked priming paradigm [22], prime stimuli are presented
briefly enough to prevent the use of conscious predictive strategies
that often contaminate standard priming effects. This paradigm is
ideally suited for investigating early perceptual processing of
familiar visual objects such as letters, without having to degrade
target stimuli, and therefore complements the psychophysical
procedures described in Box 1. In the majority of masked priming
studies of letter identification performed up to now, primes were
complete letters that varied in terms of their visual similarity with
the target (e.g. c-C versus a-A) and whether or not they were
nominally identical to the target (e.g. a-A versus c-A). The two main
tasks used in these studies were found to be differentially sensitive
to these two priming manipulations. The alphabetic decision task
(speeded classification of letters versus non-letters) was found to
be more sensitive to visual overlap than nominal overlap [51–54],
whereas the letter naming task was much more sensitive to
nominal overlap than to visual overlap [51,54,55] (Figure I). One
methodological conclusion from this research is that the letter
naming task might be overly sensitive to phonological-articulatory
factors, which are interesting in their own right, but render the task
relatively insensitive to visual factors. However, phonological
priming effects in letter naming only occur with complete
phonological overlap across primes and targets. Similar sounding
letters (P-B) do not facilitate letter naming relative to different
sounding letters (F-B) [51,55], whereas homophones of letter Figure I. Effects of nominal identity and visual similarity in three different tasks
combined with masked priming: letter naming (NAM), alphabetic decision (ADT)
targets (e.g. sea-C) facilitate letter-naming responses to the same
and perceptual identification (PIT). Effects of nominal identity are calculated by
extent as same letter primes (c-C) [51]). The absence of priming
comparing visually dissimilar same letter primes (a-A) and different letter primes
from similar sounding letters might be because the overlap is (b-A). Effects of visual similarity are calculated by subtracting the priming effect
practically always on the second phoneme, with mismatching of a visually similar same letter prime (c-C versus b-C) from the effect of a
information on the initial phoneme. Masked priming studies of visually dissimilar prime (a-A versus b-A). Net priming effects are averaged over
word and object naming have shown that initial phoneme overlap relevant conditions in two studies testing these conditions [51,54], and
is the key factor driving priming effects [56–58]. expressed as a percentage of the combined priming effect size for each task.

effects in the event-related potential (ERP) signal as a The likely neural source of this early letter-specific
function of (i) prime-target visual overlap (peaking at brain activity is left occipital-temporal cortex [26–32]. This
150 ms post-target onset); (ii) whether or not primes were fits well with research locating orthographic processing
the same letter as targets in the same case (180 ms post- associated with printed words in a small strip of left fusi-
target) and (iii) whether or not primes were the same letter form gyrus called the visual word form area (VWFA, Ref.
as targets independently of case (200 ms post-target). This [33]). Single letters dissociated from objects, faces, digits
constitutes important evidence in favor of a generic hier- and letter strings in an area situated anterior to the peak of
archical model of letter identification in which visual fea- the VWFA [27,28,32] (Figure 3). Given their more anterior
tures are mapped onto abstract letter identities via a series location, neurons in this area possibly instantiate abstract
of increasingly invariant representations. location-invariant letter detectors to be distinguished from
A comparison of ERP waveforms generated by letter and more location-specific letter detectors involved in proces-
pseudo-letter stimuli can also provide useful information sing letter strings [34,35]. Furthermore, the fact that the
about the time-course of letter identification. One study letter-specific region and the VWFA fall within a more
found that the amplitude of the N170 component is larger general object processing region [29] is in line with the
for letters than pseudo-letter (false font) stimuli [24]. neuronal recycling hypothesis of Dehaene et al. [34].
Another study using pseudo-letters that were matched
to the letter stimuli in terms of component features, found Putting the parts together
that ERP waveforms diverged as early as 145 ms post- Recent evidence in favor of a generic feature-based hier-
stimulus onset [25] (Figure 3). In this study, amplitude of archical approach to letter perception therefore provides
the P2 ERP component differed across the 14 letters that hope that a complete account of the processes involved in
were tested, but some information about individual letter identifying isolated letters is within our grasp. More soph-
identities was already available in the waveform before the isticated models, such as the one shown in Figure 4, will
peak of the P2. This was indicated by the fact that item- probably include multiple layers of simple and complex
level voltage values in this time window were found to features converging on case-specific and possibly font-
correlate significantly with predicted letter identification specific letter detectors, which map in turn onto more
latencies derived from different versions of a generic inter- abstract shape-invariant letter representations [23,34,
active-activation model of letter perception, comprising 36–38].
feature and letter detector layers [4]. The best model There are several key questions that need to be
was one with excitatory feedforward and feedback connec- addressed in future developments of hierarchical fea-
tions between layers, and within-level inhibition across ture-based models. One concerns how the spatial
letters, but no between-level (feature-letter) inhibition. relations between the different features are coded, if at

384
 Review Trends in Cognitive Sciences Vol.12 No.10

Figure 3. Examples of letter-specific brain activity. ERP results from Rey et al. [25] shown as (a) scalp maps of voltage differences obtained by subtracting ERP amplitudes
generated by letter and pseudo-letter stimuli in three different time windows, and (c) grand average waveforms for letters and pseudo-letters over electrodes shown in (b).
(d) Regions of brain activation measured by fMRI, with the single letter region [28,32] indicated by red circles (left anterior fusiform), and for comparison the visual word
form area (VWFA, Ref. [33]) indicated by blue triangles. The region in white is the conjoined activation of objects and letters found by Joseph et al. [29] (left hemisphere on
the right). Reproduced, with permission, from Ref. [29].

all (because letters could be coded as an unstructured list letter identification, using the features derived from
of features). One influential theory of object recognition empirical investigations of letter perception, augmented
posits a key role for structural representations – that is a with position-in-letter information. Thus, for example,
description of the object in terms of its component parts the letter ‘W’ could be coded as: termination upper left,
and their positional relations (such as ‘x is left of y, y is termination upper right, intersection lower left, intersec-
above z’, Ref. [39]). A promising alternative is to code for tion upper centre, intersection lower right. It is this level
the position of object parts by using object-centered coor- of sophistication that might enable feature-based models
dinates, and there is recent neurophysiological evidence to better account for patterns in the empirical data (such
in favor of such object-centered coordinate systems [40]. as letter confusions). Finally, one might also consider
This approach can be easily transposed to the domain of that such spatial relations are implicitly coded in hier-

Figure 4. A hierarchical model of letter perception. Shape and location invariance are gradually achieved via a hierarchy of increasingly complex neural processors.
Adapted, with permission, from Ref. [38].

385
Review Trends in Cognitive Sciences Vol.12 No.10

archical multi-level networks such as the one depicted in 6 Selfridge, O.G. (1959) Pandemonium: a paradigm for learning. In
Proceedings of the Symposium on Mechanisation of Thought
Figure 4.
Processes (Blake, D.V. and Uttley, A.M., eds), pp. 511–529, H.M.
Another key question is: how the kind of hierarchical Stationary Office
structure illustrated in Figure 4 could be learned by a 7 Selfridge, O.G. and Neisser, U. (1960) Pattern recognition by machine.
biological system? Could the arbitrary association between Sci. Am. 20, 60–68
lowercase ‘a’ and uppercase ‘A’ be learned without super- 8 Hubel, D.H. and Wiesel, T.N. (1962) Receptive fields, binocular
interaction and functional architecture in the cat’s visual cortex.
vision (explicit tuition)? Polk and Farah [41] proposed a
J. Physiol. 160, 106–154
solution to this problem based on the influence of common 9 Larsen, A. and Bundesen, C. (1996) A template-matching
contexts on Hebbian learning. The basic idea is that we learn pandemonium recognizes unconstrained written characters with
to associate case-specific forms of the same letter by the fact high accuracy. Mem. Cognit. 24, 136–143
that they commonly occur in the context of case invariant 10 Neisser, U. (1966) Cognitive Psychology, Appleton-Century-Crofts
11 Gibson, E.J. (1969) Principles of Perceptual Learning and Development,
letters (such as in ‘map’ and ‘MAP’, in which case-specific ‘a’ Appleton-Century-Crofts
is surrounded by case-invariant ‘m’ and ‘p’). Alternatively, or 12 Geyer, L.H. and DeWald, C.G. (1973) Feature lists and confusion
perhaps in conjunction with this unsupervised learning, matrices. Percept. Psychophys. 14, 471–482
children could learn to link ‘a’ and ‘A’ on hearing the same 13 Keren, G. and Baggen, S. (1981) Recognition models of alphanumeric
letter name associated with each form. Finally, Hinton characters. Percept. Psychophys. 29, 452–466
14 Podgorny, P. and Garner, W. (1979) Reaction time as a measure of
[42,43] has recently demonstrated that a multi-layered net- inter- and intraobject visual similarity: Letters of the alphabet.
work implementing increasingly complex features can be Percept. Psychophys. 26, 37–52
trained to recognize handwritten digits with high levels of 15 Jacobs, A.M. et al. (1989) Letter perception in peripheral vision: a
accuracy. temporal discrimination matrix using eye movements. Percept.
Psychophys. 46, 95–102
16 Courrieu, P. et al. (2004) Inverse discrimination time as a perceptual
Conclusions and future directions distance for alphabetic characters. Vis. Cogn. 11, 901–919
What is the letter ‘a’? This review article has shown that a 17 Courrieu, P. (2002) Straight monotonic embedding of data sets in
rich empirical database has accumulated over the years, Euclidean spaces. Neural Netw. 15, 1185–1196
and there is hope that this will provide the necessary 18 Pelli, D.G. et al. (2006) Feature detection and letter identification.
Vision Res. 46, 4646–4674
constraints for finding the answer to Hofstadter’s [1] ques-
19 Gosselin, F. and Schyns, P.G. (2001) Bubbles: a technique to reveal the
tion. A major step towards achieving this goal has been use of information in recognition. Vision Res. 41, 2261–2271
made in recent years. Research applying psychophysical 20 Fiset, D. et al. Features for uppercase and lowercase letter
techniques has provided convincing evidence that letters identification. Psychol. Sci. (in press)
are indeed identified via their component features, and 21 Solomon, J.A. and Pelli, D.G. (1994) The visual filter mediating letter
identification. Nature 369, 395–397
there is important new evidence concerning the precise 22 Forster, K.I. and Davis, C. (1984) Repetition priming and frequency
nature of these features. At the same time, research using attenuation in lexical access. J. Exp. Psychol. Learn. Mem. Cogn. 10, 680–
electrophysiological recordings has started to provide valu- 698
able information about the time-course of component pro- 23 Petit, J.P. et al. (2006) On the time-course of letter perception:
cesses in letter identification, and functional magnetic a masked priming ERP investigation. Psychon. Bull. Rev. 13, 674–
681
resonance imaging (fMRI) studies have begun to isolate 24 Wong, A.C.N. et al. (2005) An early electrophysiological response
the brain regions involved in letter perception (as opposed associated with expertise in letter perception. Cognitive, Affective,
to letter strings and other visual objects – Figure 3). &. Behav. Neurosci. 5, 306–318
The challenge now is to develop a computational model 25 Rey, A. et al. (2008) Testing computational models of letter perception
with item-level ERPs. Cogn. Neuropsychol. DOI: 10.1080/
of letter identification that can successfully predict empiri-
09541440802176300
cal data obtained from the wide spectrum of techniques 26 Flowers, D.L. et al. (2004) Attention to single letters activates left
that have been used to investigate letter perception up to extrastriate cortex. Neuroimage 21, 829–839
now. Future developments of these models need to be 27 James, K.H. and Gauthier, I. (2006) Letter processing automatically
constrained by behavioral, electrophysiological and brain recruits a sensory-motor brain network. Neuropsychologia 44, 2937–
2949
imaging results, and should be articulated with concurrent
28 James, K.H. et al. (2005) Letter processing in the visual system:
developments in visual object recognition and printed word different activation patterns for single letters and strings. Cognitive,
perception. It is the application of such multiple con- Affective, &. Behav. Neurosci. 5, 306–318
straints that will guarantee success in putting the right 29 Joseph, J.E. et al. (2006) fMRI correlates of cortical specialization and
parts in the right place at the right time in future compu- generalization for letter processing. Neuroimage 32, 806–820
30 Longcamp, M. et al. (2003) Visual presentation of single letters activates
tational models of letter identification. a premotor area involved in writing. Neuroimage 19, 1492–1500
References 31 Pernet, C. et al. (2005) Selective response to letter categorization
1 Hofstadter, D. (1985) Metamagical Themas: Questing for the Essence of within the left fusiform gyrus. Neuroimage 28, 738–744
Mind and Pattern, Basic Books 32 Polk, T.A. et al. (2002) Neural specialization for letter recognition.
2 Pelli, D.G. et al. (2003) The remarkable inefficiency of word recognition. J. Cogn. Neurosci. 14, 145–159
Nature 423, 752–756 33 Cohen, L. et al. (2002) Language-specific tuning of visual cortex?
3 Grainger, J. (2008) Cracking the orthographic code: an introduction. Functional properties of the visual word form area. Brain 125,
Lang. Cogn. Process. 23, 1–35 1054–1069
4 McClelland, J.L. and Rumelhart, D.E. (1981) An interactive activation 34 Dehaene, S. et al. (2005) The neural code for written words: a proposal.
model of context effects in letter perception: Part 1. An account of basic Trends Cogn. Sci. 9, 335–341
findings. Psychol. Rev. 88, 375–407 35 Grainger, J. et al. (2006) Letter position information and printed word
5 DiCarlo, J.J. and Cox, D.D. (2007) Untangling invariant object perception: the relative-position priming constraint. J. Exp. Psychol.
recognition. Trends Cogn. Sci. 11, 333–341 Hum. Percept. Perform. 32, 865–884

386
Review Trends in Cognitive Sciences Vol.12 No.10

36 Cadieu, C. et al. (2007) A model of V4 shape selectivity and invariance. 49 Majaj, N.J. et al. (2002) The role of spatial frequency channels in letter
J. Neurophysiol. 98, 1733–1750 identification. Vision Res. 42, 1165–1184
37 Fukushima, K. (2003) Neocognitron for handwritten digit recognition. 50 Sowden, P.T. and Schyns, P.G. (2006) Channel surfing in the visual
Neurocomputing 51, 161–180 brain. Trends Cogn. Sci. 10, 538–545
38 Riesenhuber, M. and Poggio, T. (1999) Hierarchical models of object 51 Bowers, J.S. et al. (1998) Orthographic, phonological, and articulatory
recognition in cortex. Nat. Neurosci. 2, 1019–1025 contributions to masked letter and word priming. J. Exp. Psychol.
39 Biederman, I. (1987) Recognition-by-components: a theory of human Hum. Percept. Perform. 24, 1705–1719
image understanding. Psychol. Rev. 94, 115–147 52 Jacobs, A.M. and Grainger, J. (1991) Automatic letter priming in an
40 Connor, C.E. et al. (2007) Transformation of shape information in the alphabetic decision task. Percept. Psychophys. 49, 43–52
ventral pathway. Curr. Opin. Neurobiol. 17, 140–147 53 Jacobs, A.M. et al. (1995) The incremental priming technique: a method
41 Polk, T.A. and Farah, M.A. (1997) Simple common contexts for determining within-condition priming effects. Percept. Psychophys.
explanation for the development of abstract letter identities. Neural 57, 1101–1110
Comput. 9, 1275–1287 54 Ziegler, J. et al. (2000) Visual and phonological codes in letter and word
42 Hinton, G.E. (2007) Learning multiple layers of representation. Trends recognition: evidence from incremental priming. Q. J. Exp. Psychol.
Cogn. Sci. 11, 428–434 53A, 671–692
43 Hinton, G.E. and Salakhutdinov, R.R. (2006) Reducing the 55 Arguin, M. and Bub, D. (1995) Priming and response selection
dimensionality of data with neural networks. Science 313, 504–507 processes in letter classification and identification tasks. J. Exp.
44 Sammon, J. (1969) A nonlinear mapping for data structure analysis. Psychol. Hum. Percept. Perform. 21, 1199–1219
IEEE Trans. Comput. 18, 401–409 56 Grainger, J. and Ferrand, L. (1996) Masked orthographic and
45 Campbell, F.W. and Robson, J.G. (1968) Application of Fourier analysis phonological priming in visual word recognition and naming: Cross-
to the visibility of gratings. J. Physiol. 197, 551–566 task comparisons. J. Mem. Lang. 35, 623–647
46 Parish, D.H. and Sperling, G. (1991) Object spatial frequencies, retinal 57 Kinoshita, S. and Woollams, A. (2002) The masked onset priming effect
spatial frequencies, noise, and the efficiency of letter discrimination. in naming: computation of phonology or speech planning. Mem. Cognit.
Vision Res. 31, 1399–1415 30, 237–245
47 Alexander, K.R. et al. (1994) Spatial-frequency characteristics of letter 58 Schiller, N.O. (2008) The masked onset priming effect in picture
identification. J. Opt. Soc. Am. A 11, 2375–2382 naming. Cognition 106, 952–962
48 Gold, J. et al. (1999) Identification of band-pass filtered letters and
faces by human and ideal observers. Vision Res. 39, 3537–3560

387