PERCEPTUAL LEARNING

HOW TRAINING INFLUENCES PERFORMANCE
Although the role of knowledge and assumptions in perception is now quite clear, the detailed ways in which past experience influences perception are less clear. Recently, experimenters have begun to examine these questions by studying how training can influence performance on apparently simple visual tasks, such as judging whether the lower line in figure 8.21 is offset to the left or right of the upper line (a vernier acuity task). [vernier acuity the ability to see very small differences in the alignment of two objects, which becomes particularly obvious when the objects are close to one another] Humans can discern the but can improve even more with practice, though this may require thousands of presentations (Fahle & Edelman, 1993). The nature of the learning can be studied by measuring the extent to which it transfers from the training stimulus to other stimuli and conditions. Thus if, after training, the vernier stimulus is rotated through 90 degrees, performance on the new task is no better than it was at the start of the experiment. Similarly, performance falls if observers are trained on one retinal location and tested on others, or trained using one eye and tested on the other.

Findings like these suggest that some of the training occurs at a site where the neurons are driven by one eye, receive input from restricted regions of the retina, and are orientation-specific. Fahle (1994) speculated that the learning might reflect changes occurring in orientation-specific neurons in V1, some of which are monocular (driven by only one eye). Others have questioned the extent and nature of the specificity of learning, and suggested that there might be a general as well as a stimulus-specific component to the observed learning effects (Beard et al., 1995). This general component might reflect, for example, a change in the ability to direct attention to particular regions of the visual field. This idea receives further support from studies into visual search conducted by Ellison and Walsh (1998). Different types of visual search not only have different behavioural characteristics, but also depend on different brain regions. So some patients with attention deficits (due to damage to the par of the brain where the temporal, parietal and occipital lobes of the brain join) may be able to perform normally on feature search [feature search visual search for a unique feature such as a particular colour or orientation (e.g. a red spot) in an array of distractors defined by different features along the same visual dimension (e.g. green spots)] tasks but are markedly impaired in conjunction search tasks (Arguin et al., 1993). Also, Ashbridge et al. (1997) used transcranial magnetic stimulation (TMS) to study the role of direction of very tiny offsets, different brain regions in visual search. In this technique, a strong magnetic field is applied briefly to the surface of a localized region of the skull, temporarily disrupting neural activity in the underlying brain region. These researchers found that stimulation of the right parietal lobe did not affect initially parallel searches, but did affect initially serial searches. Moreover, a related study found that right parietal stimulation did not affect initially serial searches once they had become parallel through training. But when the observers were switched to another task, which they initially had to perform serially, right parietal stimulation could disrupt search again (Walsh et al., 1998). Walsh et al. (1998) suggest that the right parietal lobe may be involved in setting up new templates [template an internally stored representation of an object or event in the outside world, which must be matched with the pattern of stimulation of the sensory systems before identification, recognition or naming of that object or event can occur] in the temporal lobe for processing conjunctions of, say, colour and form. Once the learning is complete, the right parietal lobe no longer plays a role in the task and so stimulating this region no longer impairs performance.

NON-VISUAL KNOWLEDGE AND PERCEPTUAL SET NON-VISUAL KNOWLEDGE AND PERCEPTUAL SET

Perceptual assumptions about lighting and noses are probably common to all humans. But there are other kinds of knowledge affecting perception which depend on linguistic, graphic and other cultural conventions. The central symbol A is perceived as ‘B’ if the vertical set of symbols is scanned, and as ‘13’ if the horizontal set of symbols in scanned. Similarly, the central letter in the two words is perceived as an ‘H’ when reading the first word, and as an ‘A’ when reading the second. Such effects depend on knowledge of a particular set of alpha-numeric conventions and of the graphology of the English language (and so would presumably not be experienced by someone who spoke and wrote only Arabic). They illustrate that non-visual knowledge can be important in visual perception. There are other situations in which the role of past experience and verbal clues become apparent.

But consider the clues ‘leaves and a Dalmation dog’, and you will probably see the dog nosing among the leaves almost instantly. Similarly, the pictures have been transformed into black blocks and black lines, so that the identity of the objects they represent may not be obvious. But again, verbal clues such as ‘elephant’, ‘aeroplane’ or ‘typewriter’ are often sufficient for the observer to identify the objects. Interestingly, once you perceive the Dalmation and the elephant, it is impossible to look at the pictures again without seeing them. These effects are sometimes described as examples of perceptual set: the verbal clues have somehow ‘set’, or programmed, the individual to interpret or perceptually organize ambiguous or impoverished stimuli in a certain way.

TRICKS OF THE LIGHT

Another powerful example of the effects of knowledge in perception . Rotating the page through 180 degrees reverses the effect. The blobs that appeared concave now appear convex, and vice versa (Ramachandran, 1995). Notice that the pattern of shading of the blobs is ambiguous. In the upper part of the figure, it could be produced if protruding blobs were illuminated from above, or if receding blobs were illuminated from below (and vice versa, for the lower half of the figure). Yet we tend to perceive them as protruding blobs illuminated from above. This is because our visual system tends to ‘assume’ (on the basis of previous probabilities) that objects in our world are lit from above (as they are in natural surroundings by our single sun), and this assumption governs the perception of ambiguous shading. Presumably, someone who lived on a plane where the only illumination came from luminous sand on the planet’s surface would see the blobs on the upper part of figure 8.16 as receding and the blobs on the lower part as protruding. If the gradient of shading is switched from vertical to horizontal, then all the blobs, whether on the top or bottom, tend to be seen as protruding. This suggests that, once the direction of illumination is clearly not vertical, it tends to be ignored. Instead, another assumption dominates perception, namely that ambiguous blobs protrude (the same assumption about vertices that governs perception of the Necker cube). Although the assumption that objects are lit from above by a single light source is important, it does not always govern our perceptions, even when it is clearly applicable. Gregory (1997) has pointed out that it may be defeated by other knowledge about very familiar objects – in particular, human faces. Gregory drew attention to the fact that the hollow mask of a face does not usually appear hollow. Instead, the receding nose appears to protrude. It is only when the mask is viewed from a short distance that stereoscopic depth information (i.e. information from both eyes) is able to overcome the ‘assumption’ that noses always protrude. What would happen if this assumption about noses were to conflict with the assumption that objects are lit from above? When the rear of a hollow mask is lit from below, the nose appears to protrude and to look as though it is lit from above, in line with both assumptions. But when the lighting is from above, the nose still appears to protrude, even though it also appears to be lit from below. Clearly the assumption that noses protrude is stronger than the assumption that objects are usually lit from above. This is probably because we have no day-to-day experience of non-protruding noses, but we occasionally experience objects lit from below by reflected or artificial light.

[Richard Gregory (1923– ) is a well-known supporter of cognitive constructionist approaches to understanding perception. Originally trained in philosophy as well as psychology, he has summarized and reviewed much experimental evidence (some of which he has provided himself ) for the ‘intelligence’ of the visual system in interpreting its input, and related this ‘top-down’ view of perception to its philosophical context. His books, especially Eye and Brain, have fired generations of students with an enthusiasm for the study of perception]

RESOLVING VISUAL AMBIGUITY

What happens when alternative probabilities are about equal? The outline (Necker) cube appears to change its orientation spontaneously. Sometimes the lower square face of the cube appears nearer, and sometimes the upper square face. This reflects the absence of depth information from shading, perspective or stereopsis (3-D vision based on differences in the visual information received by each eye) that would normally reveal the orientation of the cube. Faced with two equally good interpretations, the visual system oscillates between them. But why does our visual system fail to generate a single stable percept, which is veridical (i.e. matches the characteristics of the scene exactly), namely a flat drawing on a flat sheet of paper? The reason that the brain chooses to interpret the scene as ‘not flat’ seems to reflect the power or salience of the depth cues provided by the vertices within the figure. Further evidence for this comes from the fact that we can bias the appearance of the cube by changing our point of view. So if you fixate the vertex marked 1 in figure 8.15, the lower face will seem nearer. Fixate the vertex marked 2 and the upper face tends to appear nearer. But why does this happen? Again, the answer brings us back to probabilities. When we fixate a particular vertex, it is seen as protruding (i.e. convex) rather than receding (concave). This is probably because convex junctions are more likely in the real world. To be sure, you will see concave corners (for example, the inside corners of a room), but most concave corners are hidden at the back of an object and therefore outnumbered by convex corners at the front. You can easily test this out by simply counting how many of each type you can see from where you are sitting now.

PERCEPTION OR HALLUCINATION

Most observers perceive an inverted ‘whiter-than-white’ triangle with clearly defined edges filling the space between the black discs, each with a sector removed. This inverted triangle is illusory, since the white paper on which it is perceived is of the same luminance as that outside the triangle. It is as though, faced with the incomplete black discs and line corners, the visual system makes the best bet – that this particular configuration is likely to have arisen through an overlying object occluding complete black discs, and a complete outline triangle. In other words, since the evidence for an occluding object is so strong, the visual system creates it. A rather different example is depicted. A small spot is projected onto a large frame or screen, which is then moved. What the observer sees is the spot moving on a stationary screen. Again, this appears to reflect an assessment of relative probabilities. Small foreground objects are more likely to move than large background objects, and s this is what the observer sees.

SEEING WHAT WE KNOW

Various kinds of knowledge about the world can be shown to influence perception. One class of perceptual processes seems to reflect an assessment of what it is that particular features of the stimulus are most likely to represent.

SEEING WITHOUT KNOWING

Destruction of small parts of the cortex, for example after stroke, tumour, surgery or gunshot wounds, can result in bizarre and unexpected symptoms. Colour and motion awareness and the strange phenomenon of blindsight In the syndrome known as cerebral achromatopsia, for example, patients lose all colour sensations, so the world appears to be in shades of grey (see Sacks, 1995, for a good example, and Zeki, 1993, for historical details). If the damage is restricted to a small portion of the lower surface of the occipital lobes, the loss of colour vision can occur without any other detectable anomaly: visual acuity is normal, as are depth perception, shape understanding, and so on. Recently, another syndrome has been associated with damage to a lateral part of the occipital lobe: akinetopsia. Someone with akinetopsia loses motion awareness, so that visual stimuli all look stationary even when they are moving. These patients notice if there is a change of stimulus location, but there is no sense of pure motion ccurring between the two successive locations (Zihl et al., 1983). Syndromes like this support the theory that humans possess many specialized processing areas, as do monkeys and other primates. These specialisms contrast with the general loss of subjective vision that follows lesions of the primary visual cortex, area V1. This has been strikingly demonstrated by rare cases of damage to V1 in one hemisphere of the brain. Vision is then affected in one half of visual space, so if your right visual cortex is damaged and you look straight ahead, everything to the left of you is in some way visually absent or missing. Interestingly, though, there are some visual stimuli that can still evoke behavioural responses in the ‘blind’ half of the visual field. For example, if you hold a stick in the blind field and ask the person, ‘Am I holding this stick vertically or horizontally?’ they will say, ‘What stick? I can’t see anything over there at all.’ So you say, ‘Well, I am holding a stick, so please guess what the answ r is.’ Amazingly, these patients will answer correctly most of the time, and much more often than they would by chance guessing. Their behavioural responses to large visual stimuli, including the location, motion and orientation, presented in the blind half of the visual field will be correct more than nine times out of ten. They cannot respond to the fine details of the scene, and they cannot initiate movements towards stimuli they have not been told are there, but something remains of their previous visual capacities within the blind half of the field. This phenomenon has been termed ‘blindsight’ (Weiskrantz et al., 1974). It has been of great interest in recent studies on how subjective awareness of the visual world arises (e.g. Zeki, 1993; Weiskrantz, 1997). The ventral and dorsal streams Leading away from area V1, a distinction is generally made between two broad streams of parallel visual processing (see figures 8.10 and 8.12 above). These were initially known as the ‘what’ and the ‘where’ stream, but t ere has been some dispute over the exact role of the latter, since some researchers believe it is also involved in the visual control of movements (the ‘how’ stream), not simply in locating objects. Partly for this reason, the streams have since become known as the ‘ventral’ and ‘dorsal’ treams, emphasizing their (uncontroversial) anatomical locations, not their more controversial functional roles. The ventral stream takes mainly parvo retinal input from V1 and flows towards the inferotemporal cortex, where cells respond to the sight of whole, complex three-dimensional objects (or at least to the constellations of features that characterize these objects). Damage to this stream impairs object recognition and knowing what objects are for (Milner & Goodale, 1995; Newcombe et al., 1987). This stream includes a specialized area that deals selectively with face recognition and which is damaged in the syndrome called prosopagnosia (as in the example of the man who mistook his wife for a hat: Sack , 1985). The dorsal stream, in contrast, takes mango input and runs into the parietal lobe. It deals with locating objects and with sensorimotor coordination, mostly occurring subconsciously. Damage to the parietal lobe can hamper the ability to grasp something with the hand or post a letter through the slot in a mailbox (Milner & Goodale, 1995). With right parietal lesions particularly, it becomes difficult to recognize objects from unusual points of view (such as a bucket from above), rotate an object mentally, read a map, draw, use building blocks, and pay attention to spatial locations especially on the left side of space (a phenomenon known as spatial ‘neglect’; Robertson & Halligan, 1999). In summary, these different lines of evidence strongly support the idea of parallel processing. However, they do not explain why our behaviour is not a bundle of reflex reactions to sensory stimuli. In the next sections, we consider the role of different types of cognitive knowledge in perception.