Machine Learning Impacts the 21st Century Workforce
Machine Learning is capable of a plethora of innovations. There is not yet an agreement on the tasks where machine learning systems will excel.
Technology Briefing
|
|
|
Technology Briefing is brought to you by association with Audio-Tech, publishers of critically acclaimed programs including: Trends Magazine.
Subscribe to their monthly reports and learn about big ideas, new products, new management techniques, breakthrough concepts, and trailblazing technologies.
|
|
Transcript
Over
the past several decades, computers have transformed work in almost every
sector of the economy. Today, we are at the beginning of an even
larger and more rapid transformation due to recent advances in machine
learning. Machine learning based on neural networks is arguably the
branch of artificial intelligence that will have the most widespread near-term
commercial impact, because of its ability to accelerate the pace of automation
itself.
Like
the steam engine and electricity, MACHINE LEARNING is a general-purpose
technology, capable of spawning a plethora of additional innovations and
capabilities. However, there is not yet a
widely-shared agreement on the tasks where MACHINE LEARNING systems will excel.
Therefore, there is little agreement on the specific expected effects on the
workforce and on the economy more broadly.
Recently,
Erik Brynjolfsson of MIT and and Tom Mitchell of Carnegie Mellon addressed
these questions in a study published in the journal Science. That study, based upon an understanding
of what the current generation of MACHINE LEARNING systems can and cannot do,
examined the key implications of MACHINE LEARNING for the workforce.
It
found that while parts of many jobs may be "suitable for MACHINE LEARNING,"
other tasks within these same jobs do not fit the criteria for MACHINE LEARNING
well. Hence, its effects on employment are more complex than the simple "replacement and substitution story" emphasized by most analysts. Furthermore,
the current economic effects of MACHINE LEARNING are relatively limited; so, we
are not facing the imminent "end of work," which is
sometimes proclaimed. However, the longer-term implications for the
economy and the workforce are profound.
One
of the biggest areas of concern about machine learning is its impact on
economic inequality. To date, automation is just one factor, others
include increased globalization and mass migration. However, the
potential for large and rapid changes due to MACHINE LEARNING, suggests that
its economic effects could be disruptive, creating both winners and losers.
Ensuring a positive outcome will require considerable attention form
policy-makers, business leaders, technologists, and researchers.
Attention
to unintended consequences is important because the implications of MACHINE
LEARNING are harder to forecast than those of previous digital
technologies. Why? Because MACHINE LEARNING has potential to
address a daunting challenge that we've faced with all prior technologies.
As the philosopher Michael Polanyi observed, "we know more than we can tell." Recognizing a face, riding a bike,
and understanding speech are tasks humans know how to do very well; but our
ability to reflect on how we perform these tasks is poor. For that
reason, we cannot codify many tasks easily, or perhaps at all, using a set of
formal rules. Thus, prior to MACHINE LEARNING, "Polanyi's paradox" limited the
set of tasks that could be automated by programming computers.
Until
recently, creating a new computer program involved a labor-intensive process of
manual coding. But this expensive process is increasingly being augmented or
replaced by a more automated process of running an existing MACHINE LEARNING
algorithm on appropriate training data. Today, MACHINE LEARNING
algorithms have, in many cases, made it possible to automatically train
computer systems to be more accurate and more capable than those manually
programmed by humans.
The
importance of this shift is two-fold. First, in a growing
subset of applications, such as face recognition and credit card fraud
detection, this paradigm can produce more accurate and reliable programs than
human programmers. And,
Second, this paradigm
can dramatically lower costs for creating and maintaining new software. This
lowered cost reduces the barrier to experimenting with and exploring potential
computerization. And it encourages development of computer systems that
will automatically automate many types of routine
workflows with little or no human intervention.
Such
progress in MACHINE LEARNING has been particularly rapid in the past 6 to 8 years
due in large part to the sheer volume of training data now available for some
tasks. In many cases it is large enough to capture crucial and
previously-unnoticed regularities in the tasks. Often these applications
use data sets that are impossibly large for any person to examine or
comprehend, but which are easily within the processing ability of MACHINE
LEARNING algorithms. When large enough training data sets are available, as in
dermatology diagnosis, the game of Go, or detecting potential credit card
fraud, MACHINE LEARNING can sometimes produce computer programs that outperform
the best humans at the task.
Also
critical to MACHINE LEARNING progress has been the combination of improved
algorithms, including deep neural networks (or DNNs) and extremely fast
computer hardware. For example, Facebook switched from phrase-based
machine translation models to DNNs for more than 4.5 billion language
translations each day. DNNs for image recognition have also driven error
rates on ImageNet,
down from more than 30% in 2010 to less than 3% today. Similarly, DNNs
have helped improve error rates from 8.4% to 4.9% in voice recognition since
July 2016. -- Breaking through this 5% threshold for image and speech
recognition is important because that is roughly the error rate of human
experts when given similar data.
Although
recent advances in the capabilities of MACHINE LEARNING systems are impressive,
they are not equally suitable for all tasks. The current wave of successes
draws heavily on a paradigm known as "supervised learning," which typically
uses DNNs. This combination can be immensely powerful in domains that are
well-suited for them. However, their competence is also dramatically
narrower and more fragile than human decision-makers, and there are many tasks
for which this approach is completely ineffective.
How
can you determine whether machine-learning can effectively automate an
application? Brynjolfsson and Mitchell have identified eight
criteria to look for:
Criterion
#1: It involves learning a function that maps well-defined inputs to
well-defined outputs. These functions often include classification and prediction.
Although MACHINE LEARNING may learn to predict the Y value associated with any
given input X, this is a learned statistical correlation that might not capture
causal effects.
Criterion
#2: Large digital data-sets containing input-output pairs are
available. The
more training examples that are available, the more accurate the
learning. One of the remarkable characteristics of DNNs is that
performance in many domains does not become asymptotic after a certain number
of examples. So, it is especially important that
all of the relevant input features be captured in the training data. Although,
in principle, any arbitrary function can be represented by a DNN, computers are vulnerable to mimicking and
perpetuating unwanted biases present in the training data and to missing
regularities that involve variables that they cannot observe. Digital data sets
can often be created by
monitoring existing
processes,
hiring humans to
explicitly tagor label portions of the data, or by
simulating the
relevant problem setting.
Criterion
#3: The task provides clear feedback with clearly definable goals and
metrics. MACHINE LEARNING works
well when we can clearly describe the goals, even if we cannot necessarily
define the best process for achieving those goals. That contrasts with earlier
approaches to automation. The ability to capture the input-output decisions of
individuals, although it might allow learning to mimic those individuals, might
not lead to optimal system-wide performance because the humans themselves might
make imperfect decisions. Therefore, having clearly defined system-wide metrics
for performance (for example, to optimize traffic flow throughout a city rather
than at a particular intersection) provides a "gold standard" for the MACHINE
LEARNING system. MACHINE LEARNING is particularly powerful when training data
are labeled according to such gold standards, thereby defining the desired
goals.
Criterion
#4: The function involves no long chains of logic or reasoning that depend on
diverse background knowledge or common sense. MACHINE LEARNING systems are very strong at
learning empirical associations in data but are less effective when the task
requires long chains of reasoning or complex planning that rely on common sense
or background knowledge unknown to the computer.
Criterion
#5: The function requires no detailed explanation of how the decision was
made. Large neural nets learn
to make decisions by subtly adjusting up to several hundred million numerical
weights that interconnect their artificial neurons. Explaining the reasoning
for such decisions to humans can be difficult because DNNs often do not make
use of the same intermediate abstractions that humans do. While work is under
way on explainable AI systems, current systems are relatively weak in this
area. For example, whereas computers can diagnose certain types of cancer or
pneumonia as well as or better than expert doctors, their ability to explain
why or how they came up with the diagnosis is poor when compared with human
doctors. For many perceptual tasks, humans are also poor at explaining their
thinking, for example, how they recognize words within the sounds they hear.
Criterion
#6: The function involves tolerance for error and no need for provably correct
or optimal solutions. Nearly all MACHINE LEARNING algorithms derive their solutions
statistically and probabilistically. As a result, it is rarely possible
to train them to 100% accuracy. Even the best clinical diagnosis, speech
recognition, and object recognition systems make errors (as do the best humans).
Therefore, tolerance to errors in the learned system is an important criterion
constraining adoption.
Criterion
#7: The phenomenon or function being learned should not change rapidly over
time. In general, MACHINE
LEARNING algorithms work well only when the distribution of future test
examples is similar to the distribution of training examples. If these
distributions change over time, then retraining is typically required, and
success therefore depends on the rate of change, relative to the rate of acquisition
of new training data (for example, email spam filters do a good job of keeping
up with adversarial spammers, partly because the rate of acquisition of new
emails is high compared to the rate at which spam changes). And,
Criterion
#8: No specialized dexterity, physical skills, or mobility is
required. Robots are still quite
clumsy compared with humans when dealing with physical manipulation in
unstructured environments and tasks. This is not so much a shortcoming of
MACHINE LEARNING itself, but instead, a consequence of the state-of-the-art in
general of physical-mechanical manipulators for robots.
Prior
to MACHINE LEARNING, the primary impact of Information Technology has been on a
relatively narrow swath of routine, highly structured and repetitive tasks. This has been a key reason that labor
demand has fallen for jobs in the middle of the skill and wage spectrum, like
clerks and factory workers, whereas demand for labor at the bottom (such as
janitors or home health aides) and the top (such as physicians) has remained
strong in most advanced countries. But a much broader set of tasks will be automated or
augmented by machines over the coming years. This includes tasks for which
humans are unable to articulate a strategy, but where statistics in data reveal
regularities that define a strategy. And while a framework separating routine
tasks from nonroutine tasks did an effective job of describing tasks suitable
for the last wave of automation, the set of tasks "suitable for MACHINE
LEARNING" is often very different. Thus, simply extrapolating past trends will
be misleading, and a new framework is needed.
As
Brynjolfsson and Mitchell observe, jobs typically consist of a number of
distinct but interrelated tasks. In most cases, only some of these tasks are
likely to be suitable for MACHINE LEARNING, and they are not necessarily the
ones that were easy to automate with previous technologies. For instance,
Brynjolfsson and Mitchell found that a MACHINE LEARNING system can be trained
to help lawyers classify potentially relevant documents for a case, but would
have a much harder time interviewing potential witnesses or developing a
winning legal strategy. Similarly, MACHINE LEARNING systems have
made rapid advances in reading medical images, often outperforming humans in
some applications. However, the more unstructured task of
interacting with other doctors, and the often emotionally challenging task of
communicating with and comforting patients, are much less suitable for MACHINE
LEARNING approaches, at least as they exist today.
That
is not to say that all tasks requiring emotional intelligence are beyond the
reach of MACHINE LEARNING systems. One of the surprising implications
identified by Brynjolfsson and Mitchell is that some aspects of sales and
customer interaction are potentially a very good fit for of MACHINE LEARNING.
For instance, transcripts from large sets of online chats between sales-people
and potential customers can be used as training data for a simple chatbot that
recognizes which answers to certain common queries are most likely to lead to
sales. Companies are also using MACHINE
LEARNING to identify subtle emotions from videos of people.
Another
area where the work by Brynjolfsson and Mitchell departs from the conventional
framework is in tasks that may involve creativity. In the old computing
paradigm, each step of a process needed to be specified in advance with great
precision. There was no room for the machine to be "creative" or figure out on
its own how to solve a particular problem. But MACHINE LEARNING systems are
specifically designed to allow the machine to figure out solutions on its own,
at least for tasks that are suitable for MACHINE LEARNING. What is
required is not that the process be
defined in great detail in advance, but that the properties of the desired
solution be well-specified and that a suitable simulator exists so that the
MACHINE LEARNING system can explore the space of available alternatives and
evaluate their properties accurately. For instance, designing a complex
new device has historically been a task where humans are more capable than
machines. But so-called "generative design software" can come up with new
designs for objects that meet a set of requirements, more effectively than
anything designed by a human, and often with a very different look and feel.
This
approach works well when the final goal can be well-specified and the solutions
can be automatically evaluated as clearly right or wrong, or at least better or
worse. As a result, we can expect such tasks to be increasingly subject
to automation. At the same time, the role of humans in more clearly
defining goals will become more important, suggesting an increased role for
scientists, entrepreneurs, and those making a contribution by asking the right
questions, even if the machines are often better able to find the solutions to those questions, once they
are clearly defined.
There
are also many non-technological factors that will determine the implications of
MACHINE LEARNING for the workforce. As Brynjolfsson and Mitchell observe,
the total effect of MACHINE LEARNING on labor demand and wages is a function of
six distinct economic factors:
- Substitution. Computer systems created by MACHINE LEARNING
will directly substitute for some tasks,
replacing humans and reducing labor demand for any given level of output
- Price
elasticity. Automation
via machine learning may lower prices for tasks. This can lead to lower or
higher total spending, depending on the price elasticity of demand. For
instance, if elasticity is less than -1, then a decrease in price leads to a
more than proportional increase in the quantity purchased, and therefore total
spending will increase. By analogy, as technology reduced the price of
air travel after 1903, total spending on this type of travel increased, as did
employment in this industry.
- Complementarities. Task B may be an important, or even
indispensable, complement to task A which is automated. As the price of task A
falls, the demand for task B will increase. For example, as calculating
was automated by computers, the demand for human programmers increased. Skills
can also be complementary to other skills. For instance, interpersonal skills
are increasingly complementary to analytical skills.
- Income
elasticity. Automation
may change the total income for some individuals or the broader population. And
if "income elasticity" for a good is nonzero, this will in turn change demand
for some types of goods and the demand for the tasks needed to produce those
goods. For example, as total income has increased due to the advance of technology,
Americans have spent more of their income on restaurant meals, which could mean
more jobs for cooks and waiters.
- Elasticity
of labor supply. As wages change, the number of people working on the task will
respond. If there are many people who already have the requisite skills (for
example, driving a car for a ride-hailing service), then supply will be fairly
elastic and wages will not rise (or fall) much, if at all, even if demand
increases (or falls) a lot. In contrast, if skills are more difficult to
acquire, such as becoming a data scientist, then changes in demand will mainly
be reflected in wages, not employment. And,
- Business
process redesign. Entrepreneurs, managers, and workers constantly work to reinvent
relevant processes. When faced with new technologies, they will change the
production process, by design or through luck, and find more efficient ways to
produce output. These changes can take time and will often
economize on the most expensive inputs, increasing demand elasticity.
Similarly, over time, individuals can make a choice to respond to higher wages
in some occupations or places by investing in developing the new skills
required for work or moving to a new location, increasing the relevant supply
elasticity. With regard to machine learning, Brynjolfsson and Mitchell cite Le
Chatelier's principle, which says, "both demand and supply
elasticities will tend to be greater in the long run than in the short run as
quasi-fixed factors adjust."
It's
important to remember that machine learning won't change everything
overnight. As Brynjolfsson and Mitchell remind us, adoption and diffusion
of technologies often takes years or even decades because of the need for
changes in production processes, organizational design, business models, supply
chains, legal constraints, and even cultural expectations. Such complementary factor
are ubiquitous in modern organizations and economies; they are also subtle and
difficult to identify, and they can create considerable inertia, slowing the
implementation of radical new technologies.
Applications that require complementary
changes on many dimensions will tend to take longer to affect the economy and
workforce than those that require less redesign of existing systems. For
instance, integration of autonomous trucks onto city streets might require
changes in traffic laws, liability rules, insurance regulations, traffic flow,
and the like, whereas the switch from talking to a human assistant to a virtual
assistant in a call center might require relatively little redesign of other
aspects of the business process or customer experience.
Over
time, another factor becomes increasingly important: new goods, services,
tasks, and processes are always being invented. These inventions can lead to
the creation of altogether new tasks and jobs and thus can change the magnitudes and
direction of the relationships previously mentioned. Historically, as some
tasks have been automated, the freed-up labor has been redeployed to producing
new goods and services or undertaking new, more effective production processes.
Such innovations have been more important than increased capital, labor, or
resource inputs as a force for raising overall incomes and living standards.
MACHINE LEARNING systems may accelerate this process for many of the tasks that
fit the criteria discussed earlier, by partially automating automation itself.
Given
this trend, we offer the following forecasts for your consideration.
First, the
accelerating availability of training data, will enable us to cost-effectively
automate more and more tasks.
As
more data are generated and pooled, we will discover which tasks can be
automated cost-effectively by MACHINE LEARNING. At the same time, we will
collect data even more rapidly to create even more capable systems. And,
unlike solutions to tasks mastered by humans, many solutions to tasks automated
by MACHINE LEARNING can be disseminated almost instantly worldwide. So, there
is every reason to expect that future enterprise software systems will be
written to embed MACHINE LEARNING in every online decision task, so that the
cost of attempting to automate those tasks will come down even further.
Second, over the next
twenty years machine learning will enhance productivity in virtually every
industry, everywhere.
The
ultimate scope and scale of further advances in MACHINE LEARNING may rival or
exceed that of earlier general-purpose technologies like the internal combustion
engine or electricity. These advances not only increased productivity directly
but, more importantly, triggered waves of complementary innovations in
machines, business organization, and even the broader economy.
Individuals, businesses, and societies that made the right complementary
investments-for instance, in skills, resources, and infrastructure-thrived as a
result, whereas others not only failed to participate in the full benefits, but
in some cases were made worse off. Thus, a better understanding of the precise
applicability of each type of MACHINE LEARNING and its implications for
specific tasks is critical for understanding its likely economic impact.
And,
Third, decision-makers
and consultants will use tools such as Brynjolfsson and Mitchell's 21-point "suitability for machine-learning" framework to determine where the best
opportunities for each company and industry lie.
This
framework is reproduced in this month's printable edition of Trends. Every manager should examine it and ask how
it might be applied to their industry.
|
No comments have been submitted to date.
|
|
|
|
|
