Journal of Vision - Perceptual learning: A case for early selection, by Fahle

Perceptual learning is any relatively permanent change of perception as a result of experience. Visual learning leads to sometimes dramatic and quite fast improvements of performance in perceptual tasks, such as hyperacuity discriminations. The improvement often is very specific for the exact task trained, for the precise stimulus orientation, the stimulus position in the visual field, and the eye used during training. This specificity indicates location of the underlying changes in the nervous system at least partly on the level of the primary visual cortex. The dependence of learning on error feedback and on attention, on the other hand, proves the importance of top-down influences from higher cortical centers. In summary, perceptual learning seems to rely at least partly on changes on a relatively early level of cortical information processing (early selection), such as the primary visual cortex under the influence of top-down influences (selection and shaping). An alternative explanation based on late selection is discussed.

Perceptual learning is any relatively permanent change of perception (usually improvement as measured by changes in perceptual thresholds or brain physiology) as a result of experience. The specification “relatively permanent” distinguishes perceptual learning from sensitization (and habituation) as well as from priming, which denote more transient changes in perception. In contrast to classical conditioning, perceptual learning involves individual stimuli rather than the association of two or more stimuli, and is not restricted to one specific response, as in operant conditioning. Perceptual learning clearly is of the implicit or procedural type: it does not lead to conscious insights that can be (easily) communicated, as is the case in declarative, or factual learning. The brain circuits storing facts and events (episodes) seem to at least partially differ from those analyzing the outer world. Hence, in amnesic syndromes, scenes may be analyzed without subsequent memory (e.g., after lesions of the hippocampal formation). Perceptual learning, on the other hand, seems to change the very cortical circuits solving the perceptual task trained. In this review, I will present results suggesting that perceptual learning is (a) very specific for elementary attributes of the stimulus, such as its orientation, and (b) able to change signal processing even on the level of primary sensory cortices that were considered as “hard wired” in adults in the not too distant past.

The perceptual task employed to test perceptual learning in most of the experiments reported here is vernier acuity, a type of visual hyperacuity (Wülfing, 1892; Westheimer, 1976). In these hyperacuity tasks, even untrained observers can attain thresholds around 10 arcsec. (These are thresholds calculated according to the conventional definition, while the appropriately calculated thresholds are a factor of 2 higher, cf. Harris & Fahle, 1995). These thresholds are at least slightly below the spacing of foveal photoreceptors, and through training, they can improve by up to a factor of 5 (i.e., to 2 arcsec in especially gifted and trained observers).

Obviously, performance in these tasks is not determined primarily by the optics of the eye nor by the photoreceptor spacing, though both factors are important because they ensure that the requirements of the sampling theorem for complete, high-resolution reconstruction of the original stimulus are met (cf., Barlow, 1981; Crick, Marr, & Poggio, 1981). Performance is instead limited by the signal-to-noise ratio of the information reaching the cortex and by the precision and selectivity of cortical processing. Hyperacuity is a good choice to study learning processes in visual perception because it is a very sensitive measure based on cortical processing. Moreover, hyperacuity is not some freak ability of over-trained laboratory observers but can be achieved even without specific training, and simultaneously at many positions in the visual field (Fahle, 1991).

In a first series of experiments, we investigated the specificity of perceptual learning on low-level features of the stimuli, such as orientation, position in the visual field, and the eye used during training. Observers usually sat 2 m or further away from an analogue monitor (HP or Tektronix), controlled by a Power-Mac computer via custom-made high-speed 16-bit D/A converters with an output rate of 1 megapixel/s. Vernier stimuli consisted of thin (1 arcmin), bright greenish lines (around 50-400 cd/m²) on a dark surround. Each of the vernier elements was usually about 10-arcmin long and presented for 100-150 ms. Observers had to indicate in a binary forced-choice task without time pressure (maximum 5-s reaction time allowed) whether the lower (or left for horizontal verniers) element was offset to the right or left (respectively up or down) relative to the upper (right) element. Observers had to indicate their choice by pressing the appropriate one of two push buttons. Usually, we recorded the number of correct responses for a vernier with a fixed offset size for experiments consisting of only two sessions, whereas thresholds were measured in experiments with more sessions. A staircase procedure (PEST; Taylor & Creelman, 1967) controlled vernier offset size in these experiments where thresholds were defined as 75% correct responses. Individual sessions lasted for about 1 hr, usually consisting of 20 blocks of 80 stimulus presentations each for the experiments with fixed vernier offsets. Only one session per observer took place each day, with sessions following each other in intervals of no more than 3 days, usually on subsequent days.

It turned out that observers, on average, significantly improved performance in a standard vernier discrimination task, though there was high inter-observer variance (Fahle & Edelman, 1993). Typically, performance improved fast initially and slower after about 10-20 min of training (see Figure 1). Earlier, we had found that the improvement in vernier discrimination through learning did not generalize to a stimulus rotated by 90° (Poggio, Fahle, & Edelman, 1992). In the present experiment, by stepwise reducing the rotation of the stimulus after the first training session in six subsequent groups of observers, we found that stimulus rotation by as little as 10° was sufficient to reduce performance to baseline (i.e., there was no generalization of improvement in vernier discrimination from one stimulus to a stimulus rotated by as little as 10°: learning had to start from scratch for the new orientation) (Figure 1). Even after a stimulus rotation of 4°, performance decreased slightly and full transfer occurred only for rotations of no more than 2° (results not shown).

Figure 1. No transfer of improvement through learning after stimulus rotation by 10 deg. Eleven observers practiced vernier discriminations with a stimulus slanted by 5 deg relative to the vertical. Performance (i.e., the percentage of correct responses) improved within 1 hr of training. On the next day, a single block of presentations at the old orientation (point immediately left of vertical red line) proved that performance remained constant over night. When the stimulus was rotated by 10 deg to a slant of 5 deg in the opposite direction, performance dropped to pretraining levels (first point to the right of vertical line). The first orientation was retested at the end of the experiment. Means and SEs of 11 observers (after Fahle, 1998).

Improvement is similarly specific for position in the visual field. Eight observers practiced vernier discrimination sequentially at eight positions in the visual field, in pseudo-random order. These positions were all located on an imaginary circle around the fovea with a radius of 10° (i.e., all stimuli were presented at 10° eccentricity). Observers practiced vernier discrimination for 1 hr at each of the positions while fixation was monitored. During this time, they improved performance, on average, by 7 % (e.g., from 80% to 87% correct responses) (Figure 2).

Figure 2. Specificity of perceptual learning for the visual field position trained. Eight observers practiced vernier discriminations sequentially at 8 positions at 10º distance from the fovea. At each position, their mean performance improved during the 1 hr of training at each position by, on average, 7% (with one exception: position 4). But when proceeding to the next visual field position, performance dropped by roughly the same amount. Hence, improvement did not transfer between different visual field positions (after Fahle, Edelman & Poggio, 1995).

However, when stimuli were presented at a new visual field position, performance dropped by 7%, hence there was no transfer at all between these visual field positions. Learning, it seems, is highly specific for location in the visual field, and even these rather regularly spaced locations that had a distance of no more than (2 × 10 × π/8) ≈ 8° from their nearest neighbors did not show any sign of being able to transfer the improvement achieved during training. These results indicate that improvement through perceptual learning is also very specific for position in the visual field. A number of different perceptual learning tasks showed a similar specificity for visual field position (Dill & Fahle, 1997; cf., however, Beard et al., 1996).

In the third experiment, a new group of six observers practiced vernier discriminations monocularly, starting with either the left or the right eye patched for five sessions. Observers improved monocular performance during these first five sessions, similar to that seen in binocular improvement. At the start of the sixth session, the contralateral eye was patched instead. Improvement did not transfer to the previously patched eye; hence, learning was specific even for the particular eye trained (Figure 3). Several other groups found a similar specificity of perceptual learning for the particular eye and visual field position trained in completely different tasks (e.g., Karni & Sagi, 1991).

Figure 3. Specificity of perceptual learning for the eye used during training. Half of observers practiced vernier discrimination with the left eye patched, whereas the right eye was patched for the second half of observers. After 5 days of training for 1 h daily, the contralateral eye was patched during training. Thresholds had improved significantly over the first 5 days, but increased with an overshoot when the patch was moved to the contralateral eye (after block 22). Means and SEs of six observers (after Fahle in Fahle & Poggio, 2002).

Discriminating the orientation of short line elements obviously relies on some form of orientation discrimination. The same seems to be true for vernier discrimination (see Watt, 1984) and also for curvature detection (see Kramer & Fahle, 1996).

Eighteen observers were divided into six groups who practiced vernier, curvature, and orientation discrimination for 1 hr each in counterbalanced order. Each of the three parts of Figure 4 shows the results of three groups of six observers, each group practicing one of the three tasks for 1 hr each during each of the sessions. It is obvious that observers improved through training in each of the sessions but that the improvement did not transfer to another task (cf., Fahle, 1997). Moreover, separate analysis of the data of each of the six groups revealed no indication of transfer between any pair of tasks.

Figure 4. Upper panel. Not only orientation discrimination but also vernier offset and curvature discrimination can be based on orientation cues. Here, the discrimination is between a slant to the right versus a slant to the left. Lower panel. Performance as a function of practice in three hyperacuity tasks based on discrimination of orientation cues: vernier, orientation, and curvature discrimination (after Fahle, 1997).The same number of observers practiced each of these tasks in each of the sessions; hence, the stimulus condition is counterbalanced between sessions. In each session, observers significantly improve performance, and retain this improvement on the first block of the next session, usually on the next day (first data point to the right of the first two vertical lines). Changing to a new type of orientation-discrimination task (second points to the right of lines) decreases performance to baseline level. The very last data point tests performance for the condition of the first session. Mean results and SEMs for 18 observers. Insets give slopes and correlation factors of linear regressions through the data points of each session.

A recent study of six patients suffering from amnesic syndrome supports the hypothesis of an involvement of relatively low levels of cortical processing in perceptual learning, rather than a purely cognitive level of learning. Testing of the patients was similar to that of normal observers, apart from the fact that the intervals between the response of the patient and the next stimulus presentation were 3-4 s rather than 0.5 s as with the normal observers. Moreover, patients indicated the direction of offset verbally, and the experimenter then pushed the corresponding buttons without himself seeing the stimuli. Two of the six patients tested clearly improved performance within as few as 2 sessions with 5 blocks of 80 presentations each, tested at a 1-week interval (i.e., receiving less than half the number of stimulus presentations used to train normal observers) (Figure 5; Fahle & Daum, 2002). Another two patients improved somewhat, whereas performance of the remaining two patients stayed constant, similar to the results of about 15% of the healthy student population of around 300 observers we tested so far. Although after the 1-week gap following the first session the patients did not recollect that they had ever before participated in such an experiment (and did not remember the experimenter), the performance of the six patients as a group improved significantly as a result of training. This finding indicates that perceptual learning does not require normal function of the neuronal circuits underlying explicit or declarative learning.

Figure 5. Improvement in a vernier discrimination task in six amnesic patients (after Fahle & Daum, 2002). Six patients suffering from amnesic syndrome practiced vernier discriminations for 5 blocks with 80 presentations each and for another 5 blocks 1 week later (gap between blocks 5 and 6 indicates 1-week interval). Thresholds of two patients (RE and MH) improved dramatically as a result of training, whereas those of the remaining four patients either improved somewhat (JR and HW) or not at all (AS and HS). Hence, at least some amnesic patients are capable of perceptual learning.

Does improvement through training in perceptual tasks require attention or is it automatic, that is, based on mere stimulus presentation? A recent study reports that performance for detecting the predominant direction of dot motion improves even if this motion is not consciously perceived. Hence improvement can be independent not only of attention to the stimulus but also of conscious perception (Watanabe, Náñez, & Sasaki, 2001)!

In hyperacuity learning, on the other hand, attention certainly seems to play an important role. When two vernier stimuli are presented simultaneously, resembling a cross, and only one of the verniers is attended, offset discrimination only for this stimulus improves over the course of training. Half of observers started by indicating the offset of the horizontal vernier (up versus down; cf., Figure 6; Herzog & Fahle, 1994), whereas the other half attended to the vertical vernier and indicated its offset (left versus right). After 1 hr of training, observers’ tasks were exchanged: those initially responding to the offset of the horizontal vernier now attended to the vertical one and vice versa. Performance dropped at this transition, though the stimulus had not been changed at all, just the task was different (Figure 6): an argument against motor improvement (of accommodation or fixation; see below) as the basis of perceptual learning. On the other hand, changing the motor instructions (press left button for right offset and vice versa) yielded perfect transfer of improvement (data not shown). Hence, the mere presentation of the stimulus elements was not sufficient for improvement, but the elements had to be attended to yield improved performance, in contrast to the results with random dot kinematograms (Watanabe et al., 2001).

Figure 6. Two vernier stimuli are combined. Observers start with discriminating offset directions of either the horizontal or the vertical vernier stimulus. After 1 hr of training (in the second session), they respond to the perpendicular stimulus that had not been attended to during the first session. This switch of attention to another task decreases performance without any change of the physical stimulus (after Herzog & Fahle, 1994). One of the vernier targets is shown as a dotted line in each of the stimuli of the graph only to indicate that this vernier was not attended to. The physical stimuli were always solid lines in the experiments and did not change over the course of the experiment.

Improvement through learning in hyperacuity tasks is possible even in the absence of external error feedback (McKee & Westheimer, 1978; Fahle et al., 1995) but often is significantly faster if feedback is present. If only half of the incorrect responses are followed by an error signal (incomplete feedback), observer’s improvement is almost as fast as with complete feedback where each incorrect response leads to an error signal (Herzog & Fahle, 1997).

This finding poses difficult problems for neuronal network theories of perceptual learning based on a teacher signal, which allows the observer to classify each stimulus. With partial feedback, half of the incorrect responses would be classified as being correct, and this should strongly decrease learning but it does not. (A possible remedy would be to use only the error signals for response modification, but this method would not be able to reliably discriminate between correct versus incorrect responses.) Random feedback signals, on the other hand, without any correlation to the correctness of the response, effectively prevented improvement through training if observers assumed that they received correct error feedback (Herzog & Fahle, 1997).

Improvement is about as fast with block feedback when the percentage of correct responses is indicated after each block of 80 presentations as it is for complete trial-by-trial feedback (Herzog & Fahle, 1997). Most surprisingly, manipulating the block feedback in a way similar to the condition of random feedback, by presenting a number uncorrelated with the actual performance of the observer, also prevents improvement (Herzog & Fahle, 1997). Hence, feedback can strongly influence the speed and extent of visual learning, indicating that several top-down influences, not just attention, must play a major role in this type of learning, even if it occurs partly on early levels of cortical information processing.

Extremely high visual resolution is possible only at the very center of the visual field, subserved by the foveola. Resolution starts deteriorating at a distance of less than 1 deg from the center. So to achieve optimal performance, targets for most tasks have to fall into this very center of the fovea (but which is still more than 20-30 photoreceptor diameters wide). This may not be the case for inexperienced observers in the darkish experimental room under somewhat artificial viewing conditions: Observers may not be able to maintain a sufficiently precise fixation for the duration of the whole experiment. Similarly, for optimal retinal image quality, accommodation has to be very precise, and this may not be easy to achieve throughout a whole session for inexperienced observers. So several skeptics argued that in reality, improvement through training might be the result of motor learning: improvement of accommodation, or fixation, or both. (After all, improvement through training is based on motor improvement in many forms of procedural learning.) These skeptics continued by arguing that this motor improvement would be specific for the stimulus and eye employed and hence disappear after any change of orientation or eye used in the experiment.

A simple experiment ruled out this suspicion. Half of observers started to practice a three-dot vernier discrimination task. Here, the task was to indicate, in the usual binary forced-choice way, whether the middle dot was offset to the right or left relative to an imaginary line through the two end points. The other half of observers performed a three-dot bisection task (i.e., they had to indicate whether the middle one of three dots was closer to the upper or to the lower end point). These two stimuli are very similar to each other indeed: Thresholds for discrimination of vernier offsets are in the order of 10-15 arcsec, usually higher by a factor of around 2 for the bisection task. Hence, position of the middle point differs by about 1 photoreceptor diameter between a vernier offset to the left versus to the right and similarly between displacement up versus down. Differences in the position of the middle dot are even smaller between the stimuli for two tasks (e.g., between a middle dot displaced up and one displaced to the right) (Figure 7 and Fahle & Morgan, 1996). According to the theorem of Pythagoras, this difference between dot positions would be , if n₁, n₂ are the displacements at threshold for vernier and for bisection, respectively, or ≈ 1.4n for n₁= n₂. The difference between displacements to the left versus to the right, on the other hand, would be 2n.

Figure 7. Failure to transfer perceptual improvement between virtually identical stimuli, due to task difference (after Fahle & Morgan, 1996). Half of the observers started with a three-dot vernier discrimination task; the other half of the observers started with practicing a three-dot bisection task for about 1 hr. The transition between tasks is indicated by the thin vertical lines. The next day observers exchanged tasks. There was no transfer of improvement between the tasks, though the stimuli were virtually identical. The nearest data point to the left of the vertical line (21st block) was recorded on the second day.

Improvement does not transfer between the three-dot bisection and vernier tasks. A repeated measures analysis of variance with two within factors (block and sequence) on the individual data of all observers yielded a significant difference between the first and second condition, with lower performance during the second condition (76.8 +/– 0.84) than during the first condition (80.8 +/– 0.75). These results clearly demonstrate that transfer of improvement may fail even between virtually identical stimuli. The relevant parameter seems to be the task required from the observer, and motor components such as steady fixation or accommodation do not play an important role. Improvement of any of these factors through training with either the three-dot vernier or bisection task should not be disrupted by a position change of the middle dot smaller than that between the two stimuli discriminated during the first part of the experiment!

What could possibly be the cortical mechanisms underlying the improvement of vernier discrimination through perceptual learning? We saw above that improvements on the motor side, as are common in many forms of procedural learning, can be excluded. So we have to look for improvements on the sensory side.

Two straightforward lines of reasoning based on improved sensory processing are able to explain the specificity of perceptual learning. The first one emphasizes changes in receptive field structure of specific cortical neurons, whereas the second one emphasizes improvements in signal selection. These lines of argument are not mutually exclusive, but rather differ in the type of approach: more physiological versus more formal. Clearly, changes in input selection of any neuron lead to a change in its receptive field structure and simultaneously to changed signal selection, so the two processes are intrinsically related.

For both lines of argument, the question arises concerning the exact level of processing at which improvement is achieved. This question concerning the neuronal level of perceptual learning is, in many respects, similar to the one regarding the location, in the nervous system, of the neuronal process underlying stimulus selection based on attention. As with attentional processing, one could contrast an early selection hypothesis of perceptual learning with a late selection hypothesis (for a discussion on attention-based selection processes, see Broadbent, 1958; Deutsch & Deutsch, 1963; Johnston & Heinz, 1979).

Here I address this basic controversy regarding the cortical level on which perceptual learning operates. It has been argued that the specificity of perceptual learning indicates an early level of the underlying cortical modifications (e.g., Poggio et al., 1992), whereas this has been questioned by others. Mollon and Danilova (1996; cf., also Morgan, 1992) pointed out convincingly that learning might take place on levels beyond the primary visual cortex in spite of the high stimulus specificity. These authors argued that even though neurons on these higher levels are binocularly activated, the activations stemming from each of the two eyes might differ from each other (e.g., as a result of slight differences in the photoreceptor geometry between the two eyes). The controversy will be exemplified first in terms of signal detection theory and second in terms of physiology (i.e., receptive fields). Hence, in the following, I will discuss both psychophysical and neurophysiological findings.

By selecting those signals discriminating best between two stimuli while ignoring those that respond in a similar way to both stimuli, the decision level can benefit from a greatly improved signal-to-noise ratio and, hence, improve performance (cf., Pelli, 1985; similar processes may happen during childhood; cf., Andrews, 1964). This is to say that perceptual learning might be based, to a certain amount, on changing the weights with which individual inputs influence the overall reaction of the observer. Thus, by eliminating the influence of uninformative inputs, the amount of noise at the decision level is decreased.

Rejection (e.g., by inhibition) of the less relevant signals could occur principally on all levels of signal analysis before the decision stage. Two straightforward alternatives come to mind: (1) changes in signal processing by early selection (e.g., on a level where neurons are still monocularly activated but show orientation specificity); and (2) late selection with improvement of input selection and/or signal processing on higher processing levels and lack of transfer due to subtle changes of activation patterns elicited by, for example, presentation to different eyes (for monocular presentation).

Better and more appropriate selection and processing of input signals on an early level is the most straightforward explanation for the high stimulus specificity and the lack of transfer between similar stimulus positions, orientations, and between the eyes (Poggio et al., 1992).

But such a permanent change in input selection on an early stage would interfere with other perceptual tasks. Moreover, purely bottom-up driven modifications of input selection cannot explain dependence of performance on feedback and the lack of transfer between vernier, orientation, and curvature discrimination (see above and, e.g., Herzog & Fahle, 1998).

The second alternative, more adequate input selection and processing of signals exclusively on higher processing levels, would require that the inputs, in the case of monocular stimulus presentation, are too different to allow generalization between the eyes (Mollon & Danilova, 1996; an extreme form of this alternative would be input selection on the level of the decision process). This hypothesis allows for incorporating the effects of attention and feedback. However, the crucial assumption is that the inputs from one eye differ clearly from those of the other eye, and that the inputs used at one stimulus orientation differ from those used at a very similar orientation.

This alternative explanation has to cope with several problems, too. In healthy observers, the retinal mosaic is relatively similar in the two eyes and observers cannot, even after training, indicate which eye was stimulated during short monocular presentations (Helmholtz, 1867). At the same time, the variance of the exact pattern of neuronal activity evoked by a simple line stimulus must be enormous in each of the eyes due to fixation instability and tremor. Even under optimal fixation, observers move their eyes over an area with a side-length of more than 1 arcmin, and probably much more. Hence, the same stimulus (e.g., the three dots presented for the three-dot bisection/vernier discrimination task; Figure 7) will activate different groups of neurons, to a widely differing amount, even if presented repeatedly to the same eye. The amplitude of eye movements is larger than the (vernier) offset to be detected. Hence, it would be surprising if the higher cortical areas were able to achieve such an impressive improvement of performance on the basis of a highly variable monocular input within a few hundred stimulus presentations. On the other hand, observers are completely unable to make use of this improvement when analyzing the ensemble of inputs from the other retina when the partner eye is tested. These considerations argue against any explanation of eye-specific learning based on “labeled lines,” given the binocular nature of most neurons beyond the primary visual cortex.

This indirect argument does not safely refute late selection. But the additional evidence of changes induced by perceptual learning in early components (latency around 50 ms) of evoked potentials, especially over the occipital cortex (Fahle & Skrandies, 1994), and the results of animal experiments favor the early selection hypothesis. Moreover, I would argue that selection is most beneficial if exerted on a neuronal level as early as possible and that learning occurs at the lowest appropriate level in the visual system (cf., Karni, 1996).

The primary visual cortex represents the visual world in an ordered way with high positional precision and small receptive fields, whereas receptive fields increase on later levels of visual processing. Hence, suppression of all inputs not activated by a target is essential to isolate this target from nearby objects that could interfere with processing of the target. Suppression would yield highest improvement if exerted before the signals originating from the target converge, on higher cortical levels, with the signals evoked by other objects. Hence, I would advocate the first alternative: selection of the optimal input signals on an early level under top-down-control, combined with early modification of processing. A change of processing on this early level must occur, ideally in a task-specific way, under top-down control, changing, for example, lateral interactions between neurons. Task-specific selection of the input best suited to solve the task at hand would lead to temporary modifications of receptive field structure of at least some of the receptive fields on this level, and there exists an intimate relationship between input selection and cortical processing. We will now consider possible neuronal implementations for changes of input selection.

The specificity for orientation, position, and especially the eye trained immediately points to a specific location, in the visual system, of the neuronal changes underlying perceptual learning. Only in area 17, the primary visual cortex, exist neurons sensitive both to stimulus orientation (unlike those in the retina and the lateral geniculate nucleus [LGN]) and the eye stimulated (cf., Figure 8). The most parsimonious explanation for the results presented above is that perceptual learning relies on an improvement in processing by those neurons in area 17 best suited to discriminate between verniers offset in opposite directions, hence an early selection (see Poggio et al., 1992). If this explanation of the experimental results is correct, then perceptual learning would involve changes in connectivity between neurons already on the level of the primary visual cortex.

Figure 8. Schematic diagram of the visual system indicating the only location of orientation-specific monocular neurons. Neurons in the retina and in the lateral geniculate nucleus (LGN) have rotation-symmetric receptive fields; hence, they cannot discriminate between stimuli oriented at different angles and cannot subserve orientation-specific learning. Only some of the orientation-specific neurons in the primary visual cortex are monocularly driven, whereas neurons in higher visual projection areas are usually binocularly driven and should be unable to discriminate between separate stimulations of the two eyes. So the most parsimonious explanation for eye specificity of perceptual learning is based on the assumption that these monocular cells are involved in the learning.

This may be an implausible assumption because the primary visual cortex was considered for quite some time to be a hard-wired first stage of analysis (Marr, 1982). But more recent electrophysiological evidence points to plasticity even in the adult primary visual cortex (Gilbert & Wiesel, 1992; Eysel, Eyding, & Schweigart, 1998; Fahle & Skrandies, 1994; Godde, Leonhardt, Cords, & Dinse, 2002), supporting the hypothesis that perceptual learning may involve plasticity even of primary visual cortex. The recent electrophysiological results demonstrating plasticity of the adult primary visual cortex, therefore, fit nicely with the assumptions developed on the basis of psychophysical experiments demonstrating the specificity of perceptual learning as cited above. To conclude, physiological knowledge should no longer prevent us from speculating about plasticity of the primary visual cortex, hence from assuming modification of an early cortical level.

Assuming that perceptual learning modifies receptive fields in the primary visual cortex, as suggested by the hypothesis of early selection, what type of modification would we expect? Receptive fields of neurons in the primary visual cortex typically consist of an elongated excitatory field center determining the orientation specificity of the neuron and of inhibitory surrounds on both sides of the center. The neuron is most strongly activated by light falling on the receptive field center without extending into the inhibitory surround. The receptive fields of neurons with different orientation preferences and slightly differing receptive field positions are clearly able to discriminate between a straight vernier and an offset one, or between an offset to the left and an offset to the right (Figure 9; cf., Wilson, 1986). The precision of discrimination depends, among other factors, on the width of the receptive field center. So a straightforward and plausible hypothesis regarding the neuronal changes underlying perceptual learning with vernier stimuli would be that learning leads to permanently narrower receptive field centers and hence a narrower orientation band-width on an early level, such as the primary visual cortex as discussed (and rejected) by Herzog and Fahle (1998) (Figure 9).

Figure 9. A simple hypothesis about the neuronal basis of visual hyperacuity with vernier stimuli, postulating that improvement through training may be leading to narrower receptive field centers. Neurons in the visual cortex have receptive fields with antagonistic center-surround characteristics. Neurons are optimally activated by stimuli restricted to their receptive field centers without activating the surround. Narrowing of the field center means that the neurons are better able to discriminate between different stimulus orientations, and between offset directions. Most models use orientational mechanisms rotated by several 10s of degrees relative to the target orientation (off-center mechanisms (cf., left and right parts of figure; Findlay, 1973; Mussap & Levi, 1996; see also Morgan, 1986).

A general consideration and two specific examples argue against this hypothesis in its “feedforward” form. First, a change of receptive field structure could have consequences for the processing of virtually all visual stimuli, with potentially deleterious effects for other visual tasks, such as the detection of low-contrast stimuli. The receptive field may become too small to detect small differences in luminance. (For such reasons, Marr [1982] postulated a hard-wired [i.e., nonplastic] early level of processing.)

The first example immediately follows from this consideration. We know that the receptive fields in the primary visual cortex are similar to the ones shown in Figure 9 – all the information flowing to subsequent levels of analysis will pass through these early filters. If the receptive field width decreased as a consequence of perceptual learning, the signal-to-noise ratio and hence performance for orientation discrimination would improve. In consequence, performance for all perceptual tasks relying on orientation discrimination should ameliorate. Practicing vernier discriminations should lead to better orientation discrimination, and vice versa, and practicing of both tasks should transfer to curvature detection because for low curvatures, too, the feature used for discrimination seems to be an orientation cue (Kramer & Fahle, 1996). However, improvement through training did not transfer between these three tasks (see Figure 4).

The second example supporting the argument against permanent modifications of receptive fields as the basis of vernier learning is based on the specificity for stimulus orientation. If learning relied on the narrowing of early, orientation-sensitive receptive field centers, the improvement should transfer to similar stimulus orientations, as long as the same neurons are involved in the detection process. This consideration raises the question about the orientation bandwidths of cortical neurons. These have not been measured directly in humans, but neurons in the macaque cortex show orientation-bandwidths of between 20° and 60° (Movshon & Blakemore, 1973). That is to say that neuronal responses to an oriented bar are best for a defined (optimal) orientation and decrease to half that value if the stimulus is rotated by 10° to 30° to either side. (Neurons with small receptive fields, i.e., with best resolution for fine grating stimuli, show the most narrow orientation tuning). As we saw in Figure 1, the orientation tuning of perceptual learning is much finer than 20°. A rotation of the stimulus by 10° suffices to require completely new learning at the rotated stimulus orientation. This is a strong argument against the hypothesis that perceptual learning is based primarily on a permanent modification of receptive field structure in early visual areas. According to this hypothesis, training with a stimulus of a given orientation would lead to a narrowing of a wide range of orientation-selective receptive fields, whereas we find a very strict orientation selectivity of improvement.

In summary, perceptual learning of hyperacuity tasks is not just a permanent sharpening in the orientation tuning of the (relatively) peripheral orientation specific filters. The improvement is specific for each of the different tasks based on detection of differences in line orientation, and is highly specific for stimulus orientation far beyond the bandwidth of the cortical neurons subserving orientation discrimination. Hence, the assumption that continuously active modifications of early receptive field modifications are exclusive in achieving perceptual improvement in a strictly feed forward system lacks plausibility.

We realize that perceptual learning is unlikely to rely on the permanent modification of receptive field properties of “early” cortical neurons (e.g., by sharpening their orientation tuning). Dependence of perceptual learning on attention and on feedback add plausibility to the view that improvement cannot be based exclusively on exposure-dependent bottom-up processes permanently changing signal processing in the primary visual cortex (i.e., for all visual tasks; see Herzog & Fahle, 1994, 1998).

Moreover, it is undisputed that learning can change processing of visual information on higher or more cognitive levels of cortical information processing. Traditionally, the effects of perceptual learning have been attributed exclusively to changes on these levels. More recently, as the additional involvement of early levels became clear, the interplay between these different levels has been elaborated on by classifying different types or levels of perceptual learning (Ahissar & Hochstein, 1997). Hence, the question is not whether or not perceptual learning involves higher cortical levels (it does) but whether or not it additionally involves the primary visual cortex?

What might be the changes of receptive fields on higher levels of cortical signal processing? A number of different scenarios are possible. Basically, the neurons on higher levels of processing may use more complex features to discriminate vernier offsets to the right from those to the left. Through training, they would learn which input neurons on preceding or lower levels of cortical processing are best suited to discriminate between those two classes of neurons. Learning would consist, at least partly, in assuring a higher impact of these neurons on the discriminating neurons on the higher level and, hence, in changing the receptive fields of these neurons in a task-specific way.

However, these higher level cortical neurons tend to be binocularly activated and to possess large receptive fields. Therefore, they would be expected to be less specific for the exact stimulus parameters and be more disturbed by flanking lines at close distance to the test stimuli (but see below for a possible counter-argument based on labeled lines).

Let us summarize the findings based on the more physiological approach to answer the question concerning late versus early selection or change in receptive field properties. The high stimulus specificity of perceptual learning with lack of transfer between very similar stimulus orientations and the lack of transfer between the two eyes as well as the ability of observers to suppress nearby flanking lines through training (Spang, Herzog, Holland-Moritz, Stein, & Fahle, 2000) all support the argument in favor of plasticity involving the level of the primary visual cortex. But the dependence of learning on error feedback, the lack of generalization between tasks based on orientation discrimination, as well as theoretical considerations, argue strongly against plasticity on this early level. These considerations against an early modification of receptive fields could be invalidated by the postulate that the structure of receptive fields would be adjusted to each task by top-down influences from higher cortical areas. A possible explanation of the stimulus specificity of perceptual learning could then rely on a change of receptive field properties of low-level cortical neurons, in a task-specific way, under top-down control. Some final arguments for this proposal follow.

It is hardly at all possible to isolate, in a complex recurrent system, the level on which a change occurs by means of black-box methods such as psychophysics. In trying to resolve the controversy between early versus late learning, it will be helpful also to consider the results of neurophysiological studies. It is reassuring that the plasticity assumed, on the basis of the psychophysical results (e.g., Poggio et al., 1992), has a neuronal counterpart in the visual cortex (Gilbert & Wiesel, 1992; Fahle & Skrandies, 1994; Godde et al., 2002) and auditory cortex (e.g., Recanzone, Schreiner, & Merzenich, 1993; Menning, Roberts, & Pantev, 2000; Tremblay et al., 2001), suggesting early selection to take place.

Suppression of neuronal activation not relevant to solve the task is especially important if, for example, flanks are presented on both sides of vernier targets (cf., Spang et al., 2000). Through learning, the influence of the flanks can indeed be greatly reduced, a more difficult feat for neurons with large receptive fields. Similar reasoning would apply for neurons specialized for orientations different from the stimulus and for those representing an eye not stimulated under training conditions, making the assumption of early selection of signals through perceptual learning even more feasible. Moreover, the psychophysical results of Watanabe et al. (2002) argue strongly in favor of an involvement of early cortical levels in perceptual learning. These authors find greater improvement through training in lower level than in simultaneously trained higher level visual motion processing in a perceptual learning task. Local motion is processed at a very low level of motion processing, whereas global motion is processed at a higher level stage by spatiotemporal integration. Hence, the learning must take place on the lower processing level.

The hypothesis of early modification and selection of visual input under top-down control seems to be best suited to explain the psychophysical and electrophysiological findings on perceptual learning. The psychophysical indicators for plasticity in adult primary visual cortex agree well with the results of electrophysiological experiments in humans and animals. Both types of experiments provide the insight that indeed we have to accept the notion of plasticity in adult primary sensory cortices, because both the sum potentials over the occipital pole of human observers and receptive field properties of single neurons in primary visual cortex change as a result of training.

Visual perceptual learning leads to sometimes dramatic and relatively fast improvements of performance in perceptual tasks, such as hyperacuity discriminations. The improvement often is very specific for the exact task trained, the precise stimulus orientation, the stimulus position in the visual field, and the eye used during training. This specificity indicates location of the underlying changes in the nervous system at least partly on the level of the primary visual cortex. The dependence of learning on error feedback and on attention, on the other hand, proves the importance of top-down influences from higher cortical centers. In summary, perceptual learning seems to rely on changes on a relatively early level of cortical information processing, such as the primary visual cortex, under the influence of top-down selection and shaping influences. According to this view, the primary visual cortex is not a hard-wired filtering device, but modifies its input signals in a partly task-dependent way under top-down control. By learning, new types of processing are implemented on this early level. This conclusion is incompatible with older views of primary sensory cortices assuming lack of plasticity in adults, and is also incompatible with a strictly feedforward signal processing in the cortex, while advocating a model of information processing in a complex system with strong feedback from higher to lower levels of processing.