pattern recognition and visual word...

Pattern recognition and visual word forms

“But if any see fit not to agree with the opinions here expressed. . .still let them note the great magnitude of experiments. . .we have dug them up and demonstrated them with much pains and sleepless nights and great money expense. Enjoy them you, and if ye can, employ them for better purposes.” --William Gilbert, 1600

In other words: Even if you don’t like the theories, you have to explain the data.

Epigraph

The great “visual word form area” debate

Is the “visual word form area” specialized for visual word forms?

Larger debates:

Domain general vs. Domain specific

Organization-by-material vs. Organization-by-process

Roles of learning, expertise and evolution in shaping brain function.

Fusiform Gyrus and the “Visual Word Form Area”

Fusiform gyrus


VWFA: Hypothesized to “contains a population of neurons that, as an ensemble, is tuned to invariant stimulus properties and structural regularities of written words” (McCandliss et al., 2003)


VWFA: Hypothesized to “contains a population of neurons that, as an ensemble, is tuned to invariant stimulus properties and structural regularities of written words” (McCandliss et al., 2003)

The progressive development of the VWFA seemsclosely tied to the progression of skill, rather than beingmerely a matter of maturation. Shaywitz and colleagues[36] examined children’s fMRI responses across a range ofages (7–18 years) and reading abilities (impaired to highlyskilled). Activation levels of the VWFA and other nearbyregions were positively correlated with standardizedscores in grapheme–phoneme decoding ability, evenwhen the effects of age were taken into account. Suchfindings suggest that successful mastery of grapheme–phoneme conversion (i.e. decoding) is a critical precursor tothe development of the adult-like response properties ofthe VWFA. Behavioral evidence supports a similar linkbetween decoding and enhancements of word recognition[37]. Furthermore, training effects in reading-impairedchildren have recently been linked to changes in fMRIactivation, including posterior occipitotemporal regions inthe vicinity of the VWFA [38], as children attempt to linkletters to sounds.

Conclusions and future directionsWe started this article with a paradox: how can reading – arecent cultural invention – rely on a cerebral substratethat is tuned to the abstract properties of a class of stimulithat did not exist for most of human evolution? The

hypothesis of a progressive specialization of the left VWFAover the course of reading acquisition avoids problemsinherent to the notion of a ready-made ‘word recognitionmodule’. The similarity between reading and other formsof acquired visual expertise emphasizes that our visualsystem progressively adapts to the tasks to which we putit. Understanding this process of progressive adaptationrequires the study of the interplay between the rise ofperceptual expertise during skill development and theassociated changes in cortical function. We point out thepresence of convergent evidence that links the psychologi-cal effects that are critical to expert-level visual word formperception with the response properties of a particularcortical region that can be systematically identified andprobed.

Establishing a link between structure and function alsoraises additional questions of a kind that were notpreviously possible, because they are inherently productsof the particular structure involved. For example, whydoes this particular region become reliably specialized inthe same place for most individuals, and what pre-existingproperties of this region predispose it for this specificspecialization?

We propose that future work in this area should involvemapping the genetically determined properties of the

Box 2. Model of functional anatomy for invariant word perception

We propose a simple anatomical and functional model of the visualstagesofword reading,which follows thegeneral principles thatgovernobject recognition in the visual system of primates (Fig. I). Separatepathways in left and right hemispheres are integrated in the leftlateralized Visual Word Form Area (VWFA), which mediates betweenvisually specific input, andmoreabstract linguistic areas responsible forlexical, semantic and phonological processes. Although the preciseprojections from VWFA to systems involved in lexical, semantic and

phonological processes are currently less clearly defined, functionalareas probably include the left angular gyrus [52], left inferior frontalcortex [53], and temporal regions anterior to the VWFA [54]. Finally,ventral visual regions receive top-down attentional influences associ-ated with left and right parietal regions that are likely to affect allprocessing levels, and whose impairment might therefore lead tovarious forms of neglect dyslexia.

Fig. I. A tentative model of functional anatomical pathways involved in visual perception of words. Letter strings are first processed in ventral occipital regions (V1 toV4) contralateral to the stimuli, building up increasingly abstract visual representations (right). For stimuli in the left visual field, information is conveyed from the rightto the left hemisphere through fiber tracts in the splenium of the corpus callosum and over the posterior horns of the lateral ventricles [50]. This right hemispheremediated pathway and the direct left hemisphere pathway eventually converge in a structure within the left-hemispheric fusiform gyrus (the VWFA), where retinotopiccoding is lost. The two pathways can be differentially modulated by visuo-spatial attention associated with input from left (blue) and right (green) parietal areas.

TRENDS in Cognitive Sciences

VWF

Phonology

Anatomy and connectivity Functional properties

Visuo-spatialattention

Visuo-spatialattention

Lexico-semantic features

• Increasing receptive field size

• Increasingly abstract features

• Sequence of abstract letter identities

• Font invariant

• Location invariant

W-O-R-D

W d o

W o r d

W o r d

V1 V1

V2 V2

V4 V4

Opinion TRENDS in Cognitive Sciences Vol.7 No.7 July 2003 297

http://tics.trends.com

Response properties (McCandliss et al., 2003):

Responds reliably to letters and words.

May also respond to faces and objects. Responds more to letters than pseudo-letters

Modality-specific (doesn’t respond to spoken words)

Invariant with regard to retinal location, letter case, size and font (neural priming studies)

Insensitive to lexical properties (e.g. frequency)


Left Fusiform is Activated by Visual Word Forms (Cohen et al., 2002)

Words & Letters > Checkerboards in left, but not right, fusiform cluster.

Passive viewing of words, letter strings and checkerboards


Words > letters in left fusiform cluster.

Passive viewing of words, letter strings and checkerboards


How does the VWFA become specialized?

Written language is a recent cultural development (~5400 years ago), so can’t be evolution.

Children do not show letter/word specific activation in VWFA before learning to read.

Initial properties intrinsic to the region and its connectivity must determine its subsequent specialization for reading.

May be specialized for foveal objects, local object features, and invariance for position and size.

Expertise for reading in the fusiform gyrus (McCandliss et al., 2003)

Expertise in different visual categories (e.g. birds, cars) linked to enhanced perception of category members via more holistic processing of the stimulus, through functional re-organizaiton of visual areas.

Expertise for word reading may be similar.


Literate adults group letters together into a single perceptual unit (visual word form).

Speed of word recognition is unaffected by the number of letters for 3-6 letter words.

Suggests processed in parallel


Hypothesis: reading experience drives progressive specialization of a pre-existing inferotemporal pathway for visual object recognition.


Evidence:

Younger children do show word length effects for 3-6 letter words.

ERP data shows 10-year-olds have adult-like response to high frequency, but not low frequency, words.

Activation level of VWFA correlated with phoneme-grapheme decoding ability, controlling for age.

VWFA less active in adults with developmental dyslexia.

“The myth of the VWFA” (Price & Devlin, 2003)

Is the VWFA really specialized for word forms?

Neuropsychological evidence:

“pure alexics” usually have much larger lesions (including cuneus, calcarine sulcus and lingual gyrus in addition to fusiform)

“pure alexics” often have other perceptual problems (e.g. color naming, picture processing)


Is the VWFA really specialized for word forms?

Functional imaging evidence:

Also active when subjects name familiar objects, make manipulation responses to pictures of unfamiliar objects, name colors and perform auditory and tactile word processing tasks.

Fig. 1. Activation (in the posterior left midfusiform region) (P ! 0.001 uncorrected) for words and pictures of objects in the area that Cohen et al., (2002)call the VWFA. The peak coordinates (x " #42, y " #57, z " #15) from Cohen et al. (2002) are indicated by the dashed white cross hairs. The transverse,coronal, and sagittal cuts are also at x " #42, y " #57, z " #15. Row 1: reading aloud words and pseudowords relative to rest (in yellow) and feature

476 C.J. Price, J.T. Devlin / NeuroImage 19 (2003) 473–481

So what does the “VWFA” do then? 3 possibilities:

1. Different populations of neurons in the same region, one for VWFs and others for naming, object perception etc.

Not very neurally plausible.

Would require single-cell evidence.



2. A single cognitive function, not yet identified underlies all these multimodal responses.



3. the same population of neurons could support different cognitive processes, depending on their interactions with other cortical and subcortical areas.


Visual Perceptual Learning of Words and the VWFA Debate (Xue & Poldrack, 2007)

Korean characters presented in pairs: same/different judgment.

Scanned before and after training.

Difficulty controlled by amount of visual noise.

English word control task.

bottom and top–middle–bottom (see Figure 1A, for anexample). To control task difficulty in the visual discrim-ination task to be detailed later, special attention was paidin constructing the word pairs. First, these characterswere organized into 600 pairs in which the two charactersshared the same spatial layout and visual complexity butdiffered from each other by only one letter. Second, those600 pairs were then divided into 20 matched groups thatwere assigned to each training and test condition andcounterbalanced across subjects. Finally, to create equalnumber of ‘‘same’’ trial and ‘‘different’’ trial, we randomlychose one third of the pairs, broke them up to form the‘‘same’’ trials. As a result, each group of stimuli (60 char-acters, 30 pair) produced 40 test pairs (i.e., 20 ‘‘same’’ and20 ‘‘different’’). Another 240 three-letter English wordswere used as control in fMRI scan (Figure 1B). Followinga similar procedure, they were divided into four matchedgroups of 30 pairs, in which the two words only differedby one letter. We used the uppercase letters and a boldfont for the English words. Pilot data indicated that thisresulted in comparable performance on Korean andEnglish.

The experiment was created in Matlab (Mathworks,Sherborn, MA) using the Psychophysics Toolbox exten-sions (Brainard, 1997; Pelli, 1997). Each stimulus wasframed in a 113 ! 113-pixel window, drawn in whiteagainst a gray background. Random visual noise wasadded to the picture to manipulate the task difficulty(Figure 1C). The percentage of noise for English wordswas 28% for all the tasks, whereas that for Korean charac-ters varied from 28% to 38% across conditions. To createthe noise, we randomly chose a given percentage of pixelsin the picture and reversed their color (white to gray orvice versa). The pattern of the visual noise was differentfor each presentation of each stimulus.

Behavioral Training and Test

Subjects participated in five training sessions. The basicparadigm throughout the training was a same–different

judgment task. At the beginning of each trial, there was a400-msec fixation point followed by 200-msec blankscreen. Two characters then f lashed subsequently(100-msec duration with 200-msec interval). Subjectswere asked to judge whether the two characters wereidentical or not with a key press. The next trial started1 sec after the subjects making a response. Two groupsof stimuli (120 characters in total) were trained everysession (i.e., old items). One group of new stimuli (i.e.,new items) was added in each training session to allowseparation of repetition priming and skill learning(Poldrack, Selco, et al., 1999).

At the beginning of each training session, subjects tooka parameter setup test to find the desirable task difficultylevel. New items with four different noise levels (40 trialseach) were mixed. Based on subject’s performance, anoise level that corresponded to accuracy closest to 70%was chosen. Subjects then finished 4 blocks of 280 trials(80 old trials repeated 3 times, plus 40 new trials) withthe chosen noise level. There was a 2-min break be-tween each block.

Upon finishing all the training, a recognition task wasadministered to examine subjects’ explicit memory ofthese old items. Sixty old items and 60 new items wererandomly mixed, and subjects were to decide whetherthe character on the screen was new or old by pressingthe button. Each trial began with a fixation of 400 msec,followed by 200-msec blank screen. The character thenappeared and remained on the screen until subjectindicated a key press. If no responses were made in1900 msec, the character also disappeared. In either case,the next trial would start after a 1-sec interval.

fMRI Task

Each subject participated two fMRI scan sessions, onebefore training and one after training, separated by 6 to8 days. Two scanning runs were included in both pre-training and posttraining sessions: one with matchedparameters and one with parameters adjusted to match

Figure 1. Example of the stimuli. (A) Korean characters. The top and bottom show an example character with left–right–bottom, andtop–middle–bottom structure, respectively. The letters of the character are marked with different colors. (B) English words. All English wordsconsisted of three letters. Capital letters and bold font were chosen to increase the difficulty of recognition. (C) Korean characters presentedin varying amounts of visual noise; labels on each character indicate the percentage of noise.

Xue and Poldrack 1645

Visual Perceptual Learning of Words and the VWFA Debate (Xue & Poldrack, 2007)

At pre-training scan, both words and Korean characters strongly activated VWFA. Not sig. different.

After training, less VWFA activation for Korean characters, both with same level of noise as pre-training scan, and same level of performance (by increasing noise).

Suggest that “VWFA is neither specific to words nor sensitized by visual expertise with specific writing systems”

pattern recognition and visual word...

Documents