What AI
can and can’t do (yet) for your business PART II
Limitation 4: Generalizability of learning
Unlike the way humans learn, AI models
have difficulty carrying their experiences from one set of circumstances to
another. In effect, whatever a model has achieved for a given use case remains
applicable to that use case only. As a result, companies must repeatedly commit
resources to train yet another model, even when the use cases are very similar.
One promising response to this challenge
is transfer learning. In this approach, an AI model is trained to accomplish a
certain task and then quickly applies that learning to a similar but distinct
activity. DeepMind researchers have also shown promising results with transfer
learning in experiments in which training done in simulation is then
transferred to real robotic arms.
As transfer learning and other generalized
approaches mature, they could help organizations build new applications more
quickly and imbue existing applications with more diverse functionality. In
creating a virtual personal assistant, for example, transfer learning could
generalize user preferences in one area (such as music) to others (books). And
users are not restricted to digital natives. Transfer learning can enable an
oil-and-gas producer, for instance, to expand its use of AI algorithms trained
to provide predictive maintenance for wells to other equipment, such as
pipelines and drilling platforms. Transfer learning even has the potential to
revolutionize business intelligence: consider a data-analyzing AI tool that
understands how to optimize airline revenues and can then adapt its model to
changes in weather or local economics.
Another approach is the use of something
approximating a generalized structure that can be applied in multiple problems.
DeepMind’s AlphaZero, for example, has made use of the same structure for three
different games: it has been possible to train a new model with that
generalized structure to learn chess in a single day, and it then soundly beat
a world-champion chess program.
Finally, consider the possibilities in
emerging meta-learning techniques that attempt to automate the design of
machine-learning models. The Google Brain team, for example, uses AutoML to
automate the design of neural networks for classifying images in large-scale
data sets. These techniques now perform as well as those designed by humans. That’s
a promising development, particularly as talent continues to be in short supply
for many organizations. It’s also possible that meta-learning approaches will
surpass human capabilities and yield even better results. Importantly, however,
these techniques are still in their early days.
Limitation 5: Bias in data and algorithms
So far, we’ve focused on limitations that
could be overcome through technical solutions already in the works, some of
which we have described. Bias is a different kind of challenge. Potentially
devastating social repercussions can arise when human predilections (conscious
or unaware) are brought to bear in choosing which data points to use and which
to disregard. Furthermore, when the process and frequency of data collection
itself are uneven across groups and observed behaviors, it’s easy for problems
to arise in how algorithms analyze that data, learn, and make predictions. Negative
consequences can include misinformed recruiting decisions, misrepresented
scientific or medical prognoses, distorted financial models and
criminal-justice decisions, and misapplied (virtual) fingers on legal scales. In
many cases, these biases go unrecognized or disregarded under the veil of
“advanced data sciences,” “proprietary data and algorithms,” or “objective
analysis.”
As we deploy machine learning and AI
algorithms in new areas, there probably will be more instances in which these
issues of potential bias become baked into data sets and algorithms. Such
biases have a tendency to stay embedded because recognizing them, and taking
steps to address them, requires a deep mastery of data-science techniques, as well
as a more meta-understanding of existing social forces, including data
collection. In all, debiasing is proving to be among the most daunting
obstacles, and certainly the most socially fraught, to date.
There are now multiple research efforts
under way, as well as efforts to capture best practices, that address these
issues in academic, nonprofit, and private-sector research. It’s none too soon,
because the challenge is likely to become even more critical, and more
questions will arise. Consider, for example, the fact that many of these
learning and statistically based predictive approaches implicitly assume that
the future will be like the past. What should we do in sociocultural settings
where efforts are under way to spur change—and where making decisions based on
past behavior could inhibit progress (or, worse, build in resistance to
change)? A wide variety of leaders, including business leaders, may soon be
called upon to answer such questions.
Hitting
the moving target
Solutions to the limitations we have
described, along with the widespread commercial implementation of many of the
advances described here, could be years away. But the breathtaking range of
possibilities from AI adoption suggests that the greatest constraint for AI may
be imagination. Here are a few suggestions for leaders striving to stay ahead
of—or at least not fall too far behind—the curve:
Do your homework, get calibrated, and keep
up.
While most executives won’t need to know
the difference between convolutional and recurrent neural networks, you should
have a general familiarity with the capabilities of today’s tools, a sense of
where short-term advances are likely to occur, and a perspective on what’s
further beyond the horizon. Tap your data-science and machine-learning experts
for their knowledge, talk to some AI pioneers to get calibrated, and attend an
AI conference or two to help you get the real facts; news outlets can be
helpful, but they can also be part of the hype machine. Ongoing tracking
studies by knowledgeable practitioners, such as the AI Index (a project of the
Stanford-based One Hundred Year Study on Artificial Intelligence), are another
helpful way to keep up.
Adopt a sophisticated data strategy.
AI algorithms need assistance to unlock
the valuable insights lurking in the data your systems generate. You can help
by developing a comprehensive data strategy that focuses not only on the technology required to pool data from disparate systems
but also on data availability and acquisition, data labeling, and data
governance. Although newer techniques promise to reduce the amount of data
required for training AI algorithms, data-hungry supervised learning remains
the most prevalent technique today. And even techniques that aim to minimize
the amount of data required still need some data. So a key
part of this is fully knowing your own data points and how to leverage them.
Think laterally.
Transfer-learning techniques remain in
their infancy, but there are ways to leverage an AI solution in more than one
area. If you solve a problem such as predictive maintenance for large warehouse
equipment, can you also apply the same solution to consumer products? Can an
effective next-product-to-buy solution be used in more than one distribution
channel? Encourage business units to share knowledge that may reveal ways to
use your best AI solutions and thinking in more than one area of the company.
Be a trailblazer.
Keeping up with today’s AI technologies
and use cases is not enough to remain competitive for the long haul. Engage
your data-science staff or partner with outside experts to solve a high-impact
use case with nascent techniques, such as the ones discussed in this article,
that are poised for a breakthrough. Further, stay informed about what’s
possible and what’s available. Many machine-learning tools, data sets, and
trained models for standard applications (including speech, vision, and emotion
detection) are being made widely available. Sometimes they come in open source
and in other cases through application programming interfaces (APIs) created by
pioneering researchers and companies. Keep an eye on such possibilities to
boost your odds of staking out a first-mover or early-adopter advantage.
The promise of AI is immense, and the
technologies, tools, and processes needed to fulfill that promise haven’t fully
arrived. If you think you can let the technology develop and then be a
successful fast follower, think again. It’s very difficult to leapfrog from a
standing start, particularly when the target is moving so rapidly and you don’t
understand what AI tools can and can’t do now. With researchers and AI pioneers
poised to solve some of today’s thorniest problems, it’s time to start
understanding what is happening at the AI frontier so you can position your
organization to learn, exploit, and maybe even advance the new possibilities.
By Michael Chui, James Manyika, and Mehdi Miremadi McKinsey
Quarterly January 2018
https://www.mckinsey.com/business-functions/mckinsey-analytics/our-insights/what-ai-can-and-cant-do-yet-for-your-business?cid=other-eml-alt-mkq-mck-oth-1801&hlkid=ebe57c83f2bf4531a29b965fdb8d3493&hctky=1627601&hdpid=d3bf26bf-7f43-48fd-8f9d-9770bfb5f550
No comments:
Post a Comment