64 / Rotman Management Spring 2017
of fast mechanisms that allow algorithms to ‘learn’ from their
own performance via feedback that tracks their successes and
failures in replicating or predicting the data sets they are trained
to compress and replicate (or ‘understand’).
Despite the momentous advances in understanding the
role that feedback plays in learning, professional and higher ed-
ucation are lagging dangerously behind what is now both possi-
ble and desirable. The ‘lecture-homework-quiz-exam’ routines
that pervade higher education — whereby feedback is provided
en masse — lag student performance by a long time and are not
adaptive or personalized to the learner or to her task. As such,
current teaching practice — and the learning environment it
produces — lives in self-sufficient isolation from the findings
of learning science, deep learning science and engineering and
the neuroscience of learning regarding the impact of feedback
on skill and competence development.
Today’s feedback practices resemble those in effect 50 and
even 100 years ago — an inertia driven partly by the econom-
ics of higher learning and partly by the cultural imperviousness
of pedagogical practice to learning science and technology. But
the opportunity costs of this ‘knowing-doing gap’ are very high
and rising quickly: This gap represents both a significant drag
on the learning curve of students and an important opportunity
for disrupting the $2 trillion (2016 dollars) higher education in-
dustry. It is one that some organizations, as we will see, have
already laid the groundwork for.
Missing and Counterproductive Feedback Patterns
To understand how the current feedback landscape of higher
education fails learners by falling short of state-of-the-art
learning science, let’s return for a moment to the century-old
lecture-homework-quiz-exam routine that is the central model
for learning today.
Lectures present concepts, models, methods, heuristics,
along with their derivations thereof and applications. Home-
work problems, quizzes and exams often test for the norma-
tive or correct application of a skill or method to an unfamiliar
problem. Feedback on the exercise of the skill by the learner
is, for the most part, given by teaching assistants on problem
sets, quizzes and exams turned in by learners — in batches, and
days or weeks following the completion of the work. This is the
exact opposite of what learning science tells us about feedback
that maximizes skill development. Specifically, feedback in our
current model is.- TOO LATE: Graders typically take days or weeks to deliver
feedback to learners, in sharp contrast to the results of stud-
ies that indicate the importance of immediate feedback in
the development of skill. - TOO RARE: Feedback is infrequent relative to both the weight
it should receive vis à vis other learning activities (such as
listening and taking notes) — given its importance to learn-
ing, and especially to the learning of complex skills. An ar-
tificial neural network can ‘shatter’ — or, learn to classify —
a large data set containing lots of non-linear relationships
only if it receives profuse feedback about its performance
as it ‘learns’. Why would a real neural network be any
different? - TOO IMPERSONAL OR GENERAL: The feedback the learner re-
ceives is not adaptive to her specific patterns of thought or
behaviour. Because of the ‘economics of feedback’ in higher
education, there is little time or scope to adapt the feedback
to the learner’s specific goals and stock of existing skills,
which significantly decreases the actionability of the feed-
back for the learner. - TOO IMPRECISE: Learners usually receive ‘0 or 1’- type feed-
back (i.e. correct/not correct) relating to the degree to which
they answer a question or solve a problem as a whole — but
not on the specific pitfalls of the thinking or reasoning un-
derlying an incorrect or partially-correct answer. This makes
the feedback signal difficult to interpret as an action-guiding
and behaviour-correcting input. - TOO NOISY: Much of the feedback learners receive is heavily
dependent on the rapidly-changing and idiosyncratic bias-
es, moods, dispositions and physiological states of the grad-
ers. Different graders can disagree sharply on the quality of
Machine learning has made rapid advances in the last 10 years due to
algorithms that ‘learn’ from their own performance, via feedback.