Learning, according to Merriam-Webster , is the activity or process of gaining knowledge or skill by studying, practicing, being taught or experiencing something. In our case, the practicing of certain visual tasks can improve the skill of detecting and discriminating certain visual features.
There exist quite a number of different types of learning. There is first short-term learning that leads to short-term memory. Short-term learning enables us, for example, to memorize phone numbers that we hear until we are able to write them down. Another type of short-term learning involves visual impressions that we can store in short-term memory, for example when copying complex patterns.
In ophthalmology, we are more interested in the second type of learning, namely long-term learning and long-term memory. In long-term learning and memory, again, there exist two quite different types of learning. The first one deals with facts and events that can be described with words. This part of learning and memory is called explicit or declarative. The brain structure involved is mainly the medial temporal lobe; there we store facts and events from the past and learn about new facts and new events. The second type of long-term learning and memory cannot be communicated with words. It is called implicit or non-declarative learning and memory. Four subtypes of long-term learning and memory are generally discriminated. The first one is non-associative learning, namely habituation and sensitization. This is not really a long-term type of memory and learning because habituation and sensitization usually last only a few days or weeks. Habituation means that we react less strongly to a stimulus that has been presented several times in a row. On the contrary, sensitization means that we are reacting more strongly to a stimulus that was presented several times. Sensitization, of course, happens far less often than habituation. The second type of medium long-term learning and memory is called priming. It relies on the neocortex and means that a stimulus that we experienced may influence our behavior and reactions in ways that are mostly subconscious. Third, there is associative learning, namely classical and operant conditioning, as in the case of Pavlov’s dog. This type of learning and memory relies mainly on the amygdala and the cerebellum. Finally, there are procedural forms of long-term learning and memory, and personally I would count perceptual learning as one form of procedural learning, which relies on the striatum and the neocortex.
When defining perceptual learning, we can follow Gibson  who stated, “any relatively permanent and consistent change in the perception of a stimulus following practice or experience with this array will be considered perceptual learning.” The important points of this definition are, first, the part that perceptual learning means a relatively permanent and consistent change unlike, for example, dark adaptation. The second important point is that this improvement is the result of an active process. In the case of perceptual visual learning this improvement usually relies on training the perception and categorization of visual stimuli and often indeed very extensive training. Work by myself and others indicates that perceptual learning is not just a better use of sensory data on relatively “high” and complex levels of cortical processing, but that even early sensory and especially visual cortical areas can change their behavior as a result of training .
Fortunately, the processes on the cellular or neuronal level that underlie learning have been clarified by means of electrophysiological and biochemical investigations by Kandel and others . Today, we can be sure that plasticity in the nervous system relies on changes at the level of synapses. Synapses can learn, for example, to set transmitter free faster, to produce more transmitter or to set free additional second messengers. Moreover, neurons may produce additional synapses to influence other neurons better. While we do not have to consider these changes here in detail, it is certainly reassuring that the underlying mechanisms of perceptual learning on the cellular level have been clarified.
Perceptual learning is a very important process during early life. Newborns have a visual acuity clearly below 1/20 (0.05). The fast improvement of visual acuity over the first months and years of life is not only due to maturation processes, but mainly due to active learning through something I would call early perceptual learning. Both studies using visually evoked potentials (VEP) and behavioral measures, such as preferential looking, show fast improvement of visual acuity, and an increase of the visual field size (see Fig. 8). Perceptual learning is not only happening during childhood, but also in adults. While most of my patients see me to get reading glasses around the age of 45 years, there are a few non-myopic ones who come up to me 10 years later. These patients insist that they are able to read or at least were able to read until recently. I tend to believe them. Perceptual learning can enable you to guess the correct letters even from rather blurred images. And there are companies that sell apps, for example for smart phones, that enable people in this age range to read without reading spectacles by learning to decipher even rather blurred letters and words.
We decided to investigate perceptual learning mainly by means of a phenomenon called visual hyperacuity. This term denotes the fact that we as humans are able to detect features that are clearly below the diameter of the photoreceptor spacing even in the foveola, for example in stereovision and when reading a Vernier scale. The features that can be resolved are in the order of magnitude of 10 arcseconds, even for unexperienced observers, and down to 2 or 3 arcseconds for very experienced observers. These low thresholds, for example when deciding whether the lower element of a Vernier target is offset to the left or to the right relative to the upper one, is really amazing when we consider that photoreceptors have a diameter of around 25 arcseconds.
When Wülfing in the nineteenth century first described these low thresholds, people concluded that the anatomists had gotten it wrong when they calculated the size of photoreceptors. At this time, the size of photoreceptors had been measured and determined to be around 25 arcseconds when converted into an angular measure. So people reasoned that photoreceptors had to be much smaller than previously thought, due to the low thresholds measured by Wülfing . But the anatomists had gotten it right: photoreceptors are indeed much wider and larger than hyperacuity thresholds. Hering [49, 50] tried to resolve this paradox by postulating that the low thresholds are because Vernier stimuli extend over hundreds and thousands of photoreceptors and that the brain is able to average over these many photoreceptors. Unfortunately, this explanation was wrong as was shown by Ludvigh . When three dots are presented (almost aligned), then under optimal conditions, a lateral displacement of the middle dot can be detected for deviations that are again below 10 arcseconds, i.e. clearly below the photoreceptor diameter.
Only at the end of the twentieth century, the puzzle of hyperacuity has been resolved. The underlying cause for this amazing spatial acuity lies in the fact that our optics is not at all optimal. The retinal image even of the smallest star that is a light source almost as small as a mathematical point extends on the retina over several photoreceptors. So while one photoreceptor will usually be most strongly activated, its neighbors are somewhat less strongly activated. Then, the brain is able to calculate the position of this star with a precision far below the photoreceptor diameter by comparing the relative excitations of these neighboring photoreceptors. As a consequence, the spatial resolution to pinpoint the exact position of visual features relative to each other is mainly limited by signal-to-noise ratios, rather than by photoreceptor diameter or photoreceptor distance, as long as the conditions of Shannon’s sampling theorem are fulfilled . This theorem postulates that any signal can be completely reconstructed, as long as there are slightly more than two sampling points for the highest frequency that is part of this signal, in this case the image. And indeed, the density of foveal photoreceptors is sufficient to sample at least twice the highest frequency that can be produced by the optics of the eye, that is, more than 30 receptors per degree of visual angle. Hence, physics can show that there is no magic in these low perceptual thresholds in hyperacuity that enable us, for example, to detect a displacement between two lines at a distance of 100 km, once the offset is above 1.5 m!
Over the last decade we have performed quite a number of experiments on perceptual learning by using different hyperacuity tasks. Here, I will give the example of Vernier learning. As indicated above, we interpret our data as indicating that indeed to achieve the very highest performance, i.e. the very lowest thresholds, learning cannot be exclusively on relatively high levels of cortical processing but has to involve already on the early sensory cortical areas. Let me try to convince you that this hypothesis is correct. In the first experiment we presented Vernier stimuli to 12 observers. In six observers these Vernier stimuli were oriented horizontally, for the other six observers they were oriented vertically. Observers trained with these stimuli for 1 h and improved detection on average from around 50% to 70%. When we rotated the stimuli (the group that had trained with vertical stimuli now had to practice with horizontal stimuli and vice versa), the detection level dropped drastically, even slightly below 50%, and observers had to learn the new task, that only deviated from the previous task by stimulus orientation, completely from scratch, attaining 70% detection only after about one additional hour of training. In a control group where we did not change stimulus orientation no such drop of performance occurred.
We then repeated the experiment in a lengthier version, training observers for 5 h on five consecutive days. The thresholds improved from around 13 arcseconds to about slightly below 10 arcseconds during that time. Then, again, we changed orientation by 90 degrees so that observers who had trained with vertical Verniers now had to respond to horizontal Verniers. Thresholds increased strongly, to above 15 arcseconds, that is, even worse than in the untrained observers. This is to say that surprisingly, extensive training with one stimulus orientation improved performance for this orientation, but decreased performance for the stimuli rotated by 90 degrees. Again, performance improved over five additional hours of additional training to achieve the level attained for the first orientation only after these 5 h of training. This is to say that perceptual learning in the hyperacuity range is highly specific for stimulus orientation.
In a second experiment, we trained observers with one eye patched. Six observers started with the left eye patched while six further observers started with the right eye patched. The improvement was similar as in the experiment with stimulus rotation. After 1 h of training, the drop of performance after changing patch side was less pronounced than for the rotation of stimulus orientation, but for the companion experiment with long-term learning of 5 h per observer, we again found a strong decrease of performance when observers switched from seeing with one eye to the partner eye (even slightly) below the level of untrained observers. Please note that for all of these experiments, new observers were recruited for each new experiment.
These results and additional results we obtained, for example, by using visual evoked potentials that showed significant change as a result of perceptual learning already over the occipital pole , lead us to the conclusion already mentioned above, that perceptual learning can change processing already on a very early level of cortical computation before the the inputs from the two eyes are combined. If perceptual learning improved performance through better evaluation of sensory signals on higher levels of cortical processing, one would have to expect that improvements generalize from one eye to the other. One has to keep in mind that, due to tremor and small eye movements, stimuli will fall on different parts of the retina over the course of the experiment. Different parts of the same retina will differ from each other as much or maybe even more than corresponding parts of both eyes. Hence, an improvement that is specific for one eye strongly suggests that this improvement is mediated on very early levels of visual information processing that are still monocularly activated. This is to say that the old view of a hard-wired early visual cortex, as proposed for example by Marr and colleagues , does no longer hold true. Quite to the contrary, the early sensory cortical areas seem to keep some plasticity even in adults.
This has consequences not only for the therapy of amblyopia, but also for stroke patients. As long as signals reach the visual cortex, learning and compensatory mechanisms are able to improve perception and discrimination of objects. The essential condition to keep in mind is that signals from the retina have to arrive at the brain. Phenomena such as blind-sight seem to indicate that these signals do not necessarily have to arrive in the primary visual cortex, but other parts of the cortex may also be able to subserve some type of rudimentary vision. If, on the other hand, fibers are destroyed, as is the case in glaucoma or strokes on the level of the thalamus, then the resulting visual field defects cannot be made to disappear by means of perceptual learning. Training can improve the way that the visual cortex analyzes and categorizes visual stimuli, but can never compensate absolute visual field defects caused by lesions on very early levels of the visual system.
To conclude, we find that there are a number of different forms of learning and have reminded ourselves that learning dramatically improves seeing in infants and can improve visual perception at least slightly in patients and in presbyopes. We also find that in several so-called hyperacuity tasks, such as Vernier acuity and stereopsis, observers achieve spatial resolution far below the photoreceptor diameter and photoreceptor spacing even in the foveola and can thus, at least after extensive training, attain thresholds that are far below photoreceptor diameters. But improvement in perceptual learning seems under most conditions to be very specific for the exact task trained and therefore indicative of changes that involve even the level of early sensory cortical areas. Extensive research is presently under way to find training procedures leading to perceptual learning that generalizes to new tasks.