Despite the complicated assessment of an ever changing environment, observers can learn
to interpret incoming sensory data in novel environments (Scarfe & Glennerster, 2014).
Performance feedback is information that notifies a learner about their performance, and
can be generated internally or be provided by an external source (Fahle, 2002). Using
performance feedback allows an organism to quickly establish the appropriateness of their
learning strategy, and thus, should be an advantage for those organisms who were able to
do so efficiently (Shibata et al., 2009). In psychophysical experiments this often takes the
form of an auditory sound that is presented after a trial to indicate a correct or incorrect
response. Behavioural perceptual learning studies have shown that using external feedback
can improve learning and increase efficiency (Herzog & Fahle, 1999).
The role of feedback in perceptual learning is still unclear. Several studies have found that
factor (Seitz et al., 2006), but there are also those that find learning occurs without external
feedback (Liu et al., 2012a; Petrov et al., 2005; Vaina et al., 1998). This results in a
complicated pattern of empirical findings. However, feedback is not usually the primary area
of interest in perceptual learning studies, and studies that fail to find significant results are
less likely to have been published (Seitz et al., 2006).
In an early attempt to clarify the role of feedback in learning, Herzog and Fahle (1997)
tested several models using a Vernier acuity task. The first group received trial-by-trial
feedback, and all observers bar one displayed significant improvement. A second group
trained without feedback, and provided a mixture of results, including a non-significant im-
provement, no improvement and some worsening. A third block of observers were provided
with block feedback, in terms of a percentage correct after each block. Improvements were
similar to those found for individuals receiving trial-by-trial feedback. Since feedback was
not attached to an individual stimulus, they argue this type of feedback could not act as a
teaching signal. A further group were provided with random trial-by-trial feedback. There
was no learning and the authors concluded that uncorrelated feedback prevents learning.
A group of observers receiving partial feedback, on 50% of their incorrect responses, did
improve, but less than those with full feedback. Additionally, in a group receiving reverse
feedback, all but one observer adapted. Herzog and Fahle (1997) concluded that correct
feedback improves both the speed of learning and overall improvement in performance. In
the no feedback condition, results were highly variable among observers, and, on average,
no learning was found without feedback. Feedback played a role in reducing the variation
play in learning that cannot be explained as a teaching signal, since block feedback had no
signal to individual stimuli, yet learning occurred at the same rate as trial by trial feedback.
Furthermore, they note that learning with feedback was significantly easier than without it,
and there is a positive effect of feedback, especially since manipulated feedback prevented
learning. Finally, they note that learning, exclusively as a result of exposure to a stimulus, is
implausible.
However, Shibata et al. (2009) reported results that contradict Herzog & Fahle’s conclu-
sions. In a total of nine groups they manipulated the accuracy of feedback for six of the nine
groups. Manipulated groups received predefined feedback irrelevant of their actual accuracy.
For the three remaining groups; one group had unmanipulated, accurate feedback, one group
no feedback, and one group had no feedback, but were asked to judge their own level of per-
formance. The task was a same or different discrimination task, of complex gratings. Shibata
et al. (2009) defined a learning regression line for the average accuracy across subjects in
the unmanipulated accurate feedback, condition, manipulated feedback was constructed by
manipulating the gradient of the regression line, where participants’ predefined feedback
would either be the same, larger or smaller than the original. Their results revealed that
learning occurred in the group that had no feedback, which was no different to performance
in the unmanipulated accurate feedback condition. Observers who had manipulated feedback
with a larger gradient, were facilitated in comparison to the unmanipulated accurate feedback
group. Observers who received smaller gradient feedback improved at the same level as the
Vaina et al. (1995) differentiate between perceptual learning over several days, and
‘rapid’ learning, which stabilises within the first 200 repetitions. Observers tested daily on
a global motion task, with a signal of 25% for three days without feedback, and repeated
the experiment again 10 days later. They found learning occurred extremely quickly within
session for a global motion task, where observers went from chance performance to 100%
correct across the 10 days. However, this improvement was lost when direction was changed
from left/right to up/down, suggesting that improvements were specific to trained direction.
Moving away from rapid learning, Petrov et al. (2006a) suggested that feedback may
be useful when the stimulus is difficult to detect or discriminate, where it may increase
confidence and make learning more efficient. Liu et al. (2012a) predicted that there was an
interaction between accuracy and feedback. Where accuracy is high for a sufficient number
of trials, Hebbian learning predicts a high chance of learning, however, when accuracy is low
Hebbian learning alone is erratic. Alternatively, when feedback (trial-by-trial) is provided,
there should be less reliance on performance accuracy. They tested this prediction in a 6 day
contrast sensitivity (with noise) paradigm using a staircase method. Observers were divided
into high and low accuracy training groups, half of whom received trial-by-trial feedback,
and half of whom did not. They found an interaction between feedback and accuracy; when
accuracy was high, external feedback was not critical, but it was crucial when accuracy was
low. Furthermore, Liu et al. (2012a) replicated the study, finding that by mixing high and low
accuracy trials learning also occurred without the need for external feedback.
Seitz et al. (2006) also investigated if including easy exemplars could foster perceptual
discrimination task with low contrast dots, and secondly an orientation discrimination
(masked with noise) task using off cardinal (obliquely) oriented bars. While both groups
receiving external reinforcement displayed perceptual learning effects, those that experienced
no feedback failed to show learning. They concluded that internal reinforcment was not
enough to generate reinforcement signals.
Since the role of feedback is not a routinely tested paradigm, with only a handful of
studies that explicitly test for differences in learning as a result of feedback (Seitz et al.,
2006), the role of external feedback in perceptual learning remains unclear (Seitz et al., 2006,
p.972).