Christopher Steiner
in Automate This: How algorithms came to rule the world describes how
algorithms have pretty much got everywhere these days. Use the web, you’re
using algorithms. Engage with the financial world, you’re engaging with a world
driven by algorithms. Look at any image on any screen, use your mobile, satnav
and any other piece of technology and you’re in the world of algorithms. From
markets to medicine, the 21st century is being shaped by the power of
algorithms to do things faster, cheaper and better than we humans, and their
reach is getting bigger every day.
Learning
is algorithmic
Our brains clearly
take inputs but it is the processing, especially deep processing, that shapes
understanding, memory and recall. The brain is not a vast warehouse that stores
and recalls memories like video sequences and text, it is a complex,
algorithmic organ. However, it would be a mistake to see the brain as a
‘simple’ set of algorithms for learning. Take ‘memory’ alone. We have many
types of memory; sensory memory, working memory, long-term memory, flash-bulb
memory, memory for faces, and remarkably, memory for remembering future events.
Huge gains have been made in understanding how these different types of memory
work in terms of our computational brain. Even if one rejects the computational
model of the brain, alternative models do seem to require algorithmic
explanations, namely inputs, processing and outputs. Learning is therefore
always algorithmic.
You’d also be
mistaken if you thought that the world of online learning has been immune from
this algorithmic reach. From informal to formal online learning, from search to
sims, efficient online learning has already been heavily powered by algorithms.
Every time you
search for something on Google you’re using seriously powerful algorithms to
crawl the web and match its findings to your search item, then produce a set of
probable results. It’s algorithms do many other things you may not be aware of,
such as identify cheats who create lots of artificial links and keywords to
boost their ratings. Every time you use Facebook algorithms determine your
newsfeed. The more you engage with someone, the more likely you are to see
their posts, and the more Facebook people in general that engage with posts,
determines how often that post is presented to others. Every time you use
Amazon to see book rankings or get recommendations or offers, they use what
they know about you and others to offer you choices. Wouldn’t it be astonishing
if this did not happen more and more in online learning?
Even in formal
learning, there are algorithms galore. In flight simulators, it’s not just the
image-rendering and motion modelling that’s driven by algorithms, it’s also the
learning through scenarios and feedback. In games-based learning there are
algorithms aplenty. I first implemented algorithms in a BT simulator in the
1980s. What’s new is the fact that the restraints we had then have all but
disappeared, with faster processing, better memory and ability to gather large
amounts of data for use by adaptive algorithms. They’re here and they’re here
to stay.
Teaching is
essentially algorithmic, as the teacher (agent) gathers data about the
environment (knowledge and students) and attempts to deliver recommendations or
responses, based on their experience. In some cases, such as search and
citation-driven research, algorithms do this much faster and better than
humans. Teachers, tutors and lecturers, most of whom teach many students, have
limited abilities when it comes to gathering data about their students, so no
matter how good they are at teaching, tailored, personalised feedback and
learning is difficult. It is also difficult to identify what students find
difficult to learn. What, for example, are the most difficult concepts or
weaknesses in the course? Again, data gathered from individual students and
large numbers of students, allow algorithms to do their magic.
The bottom line, and
there is a bottom line here, is that warm-body teachers are expensive to train,
expensive to pay, expensive to retire, variable in quality and, most important
of all, non-scalable. This is why education continues to be so expensive. There
is no way that current supply can meet future demand at a reasonable cost. The
solution to this problem cannot be throwing money at non-scalable solutions.
Smart, online solutions that take some elements of good teaching, and replicate
them in software, must be considered. As Sebastian Thrun says, ““For me, scale
has always been a fascination—how to make something small large. I think that’s
often where problems lie in society—take a good idea and make it scale to many
people.”
We have seen how
smart algorithms, in Google. Facebook and Amazon can significantly enhance
users’ experiences, so how can they enhance the learners’ experiences? It is
often thought that data is the key factor in this process but data is inert and
only becomes useful as input to algorithms that interpret that data to produce
useful actions and recommendations. The quality and quantity of data is
important but it is the quality of the algorithms that really counts.
Smart software,
formula-based algorithms, based on the science of learning, can be used to take
inputs and make best guesses based on calculated probabilities. Multiple,
linked algorithms can be even smarter. They may use knowledge of you as a
learner, such as your past learning experiences and so on. Then, dynamically
track how you’re getting on through formative assessment, how long you’re
taking in a task, when you get stuck, keystroke patterns and so on. Add to this
data gathered from other groups of similar learners and the system gets very
smart. So smart, it can serve up optimised, learning experiences that fit your
exact needs, rather than the next step in a linear course. The fact that it
also learns as it goes make it even smarter.
Iceberg
principle
When we learn, most
of the processes going on in our brain are invisible, namely the deep
processing involved in memory and recall. This activity, that lies beneath
consciousness, is where most of the consolidation takes place. Similarly, it is
sometimes difficult to see adaptive, algorithmic learning in practice, as most
of the heavy lifting is done by algorithms that lie behind the content. Like an
iceberg the hard work is done invisibly, below the level of visible content.
Different students may proceed through different routes on their learning
journey, some may finish faster, others get help more often and take longer.
Overall, adaptive, algorithmic systems promise to take learners through
tailored learning journeys. This is important, as the major reasons for learner
demotivation and drop-out are difficulties with the course and linear, one
size-fits-all courses that leave a large percentage of students behind when
they get stuck or lost.
One symptom of the
recognition of the need for adaptive, algorithmic learning is a recent poll by Inside
Higher Ed and Gallup of college and university presidents, who saw more
potential in ‘adaptive learning’ (66%) to make a “positive impact on higher
education” than they did MOOCs (42%). The 2013 Insider Higher Ed Survey of
College and University Presidents. Conducted by Gallup. Scott Jaschik and Doug
Lederman, eds. March 2013.
Why would this be
so? Well, it’s not only college and University Presidents, the Gates Foundation
have also backed ‘adaptive learning’ as an investment. Read their two recent
commissioned reports, as well as their funded activity in this area. One UK
company, singled out by these reports, as a world class adaptive, algorithmic
learning company is CogBooks, who have already taken adaptive learning to the
US market. The reason is that it is seen as a way to tackle the problem of
drop-out and extended, expensive courses. Personalised learning, tailored to
every individual student, has come of age and may just be a solution to these
problems.
No comments:
Post a Comment