Towards a physiology of reading (1)

Listen to audio

The existence of a brain area specialized for reading, with a reproducible location from one individual to another, raises the question of its origin. Since reading is far too recent an activity to have exerted selective pressure on the evolution of our brains, we need to find out what precursors may have led to the specialization of this region in primates, which would then be "recycled" for reading in the human species.

In the macaque monkey, the ventral occipitotemporal region plays an essential role in the invariant recognition of objects and faces (Rolls, 2000). This region already achieves invariance for position, size and viewing angle: it is not interested in the position of objects, but its lesion selectively interferes with recognition of their identity. Unitary recording shows that it comprises a mosaic of neurons specialized for object shapes, and invariant for the visual cues that compose them. These neurons show great plasticity: when the animal learns to recognize novel objects, including abstract fractal shapes, many inferotemporal neurons begin to respond selectively to these shapes. When the animal matches two arbitrary shapes, these two shapes can be encoded by one and the same neuron - a neuronal pairing perhaps relevant to the learning of upper- and lower-case letter pairs.

The work of Tanaka et al (1996) has clarified the neural architecture underlying this visual pattern recognition. When an inferotemporal neuron responds to a complex shape, such as a cat, it is often possible to identify a simpler shape that activates it just as much - for example, two superimposed disks. The inferotemporal region thus comprises an "alphabet" of neurons that respond to elementary shapes, some of which evoke the shape of our letters. These shapes may have been chosen for their "non-accidental" character (Biederman, 1987), invariant to many spatial transformations, as well as for their Combinatorics capacity to represent multiple novel objects, such as the shape of a fire extinguisher (work by Tanifuji et al.) .

Visual pattern recognition thus seems to involve a hierarchical pyramid of neurons in the ventral occipitotemporal pathway, sensitive to visual units of increasing size and invariance. This architecture prefigures that of reading. It is tempting to speculate that learning to read partially modifies the preferences of certain neurons, which then begin to encode letter shapes and their combinations. An explicit model of this neural recycling has been proposed (Dehaene, Cohen, Sigman, & Vinckier, 2005). It postulates that learning to read sets up a hierarchy of local combination detectors ranging from isolated features to words, via letters, bigrams and morphemes.

It is not yet possible to test these hypotheses by recording single neurons in humans, but functional MRI has begun to test the anatomo-functional organization predicted by the model by demonstrating a topographical and hierarchical organization within the left ventral occipitotemporal pathway (Vinckier et al., 2007). Visual degradation of words, for example by character spacing, also produced results in line with our expectations. When spacing exceeds two characters, a value beyond which bigram neurons must cease to respond, a threshold of degradation is crossed and reading becomes slow, serial, and dependent on a second brain system, the occipito-parietal pathway associated with sequential orientation of visuo-spatial attention.

The interest of the psycho-physiological approach to reading proposed here is to build some bridges between brain organization and behavioral observations of reading. From this point of view, it is interesting to compare the proposed neural model with the numerous psychological data concerning the visual similarity of words. According to most cognitive models of reading, visual input is rapidly recoded into an array of letters. However, this code is not invariant to spatial movement. Other schemes have been proposed: alignment of letters to the left, to the right, centering, triplets of letters ("wickelgraphs"). None of these models, however, provides an explicit mechanism for spatial invariance, nor do they correctly predict measures of similarity between words. Various masked priming experiments indicate that the prior presentation of a fraction of letters facilitates word recognition insofar as the order of the letters is preserved (the string "jrdn" primes "jardin") or minimally altered ("bagde" primes "badge"). Jonathan Grainger and colleagues have shown that all these results can be explained by "open bigram" coding, in which the word "badge", for example, would be coded by the list of its letter pairs BA, BD, BG, BE, AD, AG, AE, DG, DE and GE (Grainger, Granier, Farioli, Van Assche, & van Heuven, 2006) - a code very close to the one we postulated by examining the neurophysiological precursors of reading. This very active field of research therefore offers hope of a convergence between neuroscience and psychology.