Despite practical challenges, we are hopeful that REVIEW

informed discussions among policy-makers and the

public about data and the capabilities of machine
learning, will lead to insightful designs of programs
and policies that can balance the goals of protecting
Machine learning: Trends,
privacy and ensuring fairness with those of reaping
the benefits to scientific research and to individual perspectives, and prospects
and public health. Our commitments to privacy and
fairness are evergreen, but our policy choices must M. I. Jordan1* and T. M. Mitchell2*
adapt to advance them, and support new tech-
niques for deepening our knowledge. Machine learning addresses the question of how to build computers that improve
automatically through experience. It is one of today’s most rapidly growing technical fields,
RE FE RENCES AND N OT ES
lying at the intersection of computer science and statistics, and at the core of artificial
1. M. De Choudhury, S. Counts, E. Horvitz, A. Hoff, in Proceedings
of International Conference on Weblogs and Social Media intelligence and data science. Recent progress in machine learning has been driven both by
[Association for the Advancement of Artificial Intelligence the development of new learning algorithms and theory and by the ongoing explosion in the
(AAAI), Palo Alto, CA, 2014]. availability of online data and low-cost computation. The adoption of data-intensive
2. J. S. Brownstein, C. C. Freifeld, L. C. Madoff, N. Engl. J. Med.
machine-learning methods can be found throughout science, technology and commerce,
360, 2153–2155 (2009).
3. G. Eysenbach, J. Med. Internet Res. 11, e11 (2009). leading to more evidence-based decision-making across many walks of life, including
4. D. A. Broniatowski, M. J. Paul, M. Dredze, PLOS ONE 8, e83672 health care, manufacturing, education, financial modeling, policing, and marketing.

M
(2013).
5. A. Sadilek, H. Kautz, V. Silenzio, in Proceedings of the
achine learning is a discipline focused ance when executing some task, through some
Twenty-Sixth AAAI Conference on Artificial Intelligence
(AAAI, Palo Alto, CA, 2012). on two interrelated questions: How can type of training experience. For example, in learn-
6. M. De Choudhury, S. Counts, E. Horvitz, in Proceedings of the one construct computer systems that auto- ing to detect credit-card fraud, the task is to as-
SIGCHI Conference on Human Factors in Computing Systems matically improve through experience? sign a label of “fraud” or “not fraud” to any given
(Association for Computing Machinery, New York, 2013),
and What are the fundamental statistical- credit-card transaction. The performance metric
pp. 3267–3276.
7. R. W. White, R. Harpaz, N. H. Shah, W. DuMouchel, E. Horvitz,
computational-information-theoretic laws that to be improved might be the accuracy of this
Clin. Pharmacol. Ther. 96, 239–246 (2014). govern all learning systems, including computers, fraud classifier, and the training experience might
8. Samaritans Radar; www.samaritans.org/how-we-can-help-you/ humans, and organizations? The study of machine consist of a collection of historical credit-card
supporting-someone-online/samaritans-radar. learning is important both for addressing these transactions, each labeled in retrospect as fraud-
9. Shut down Samaritans Radar; http://bit.ly/Samaritans-after.
fundamental scientific and engineering ques- ulent or not. Alternatively, one might define a
10. U.S. Equal Employment Opportunity Commission (EEOC), 29
Code of Federal Regulations (C.F.R.), 1630.2 (g) (2013). tions and for the highly practical computer soft- different performance metric that assigns a higher
11. EEOC, 29 CFR 1635.3 (c) (2013). ware it has produced and fielded across many penalty when “fraud” is labeled “not fraud” than
12. M. A. Rothstein, J. Law Med. Ethics 36, 837–840 (2008). applications. when “not fraud” is incorrectly labeled “fraud.”
13. Executive Office of the President, Big Data: Seizing Machine learning has progressed dramati- One might also define a different type of training
Opportunities, Preserving Values (White House, Washington,
DC, 2014); http://1.usa.gov/1TSOhiG.
cally over the past two decades, from laboratory experience—for example, by including unlab-
14. Letter from Maneesha Mithal, FTC, to Reed Freeman, Morrison, curiosity to a practical technology in widespread eled credit-card transactions along with labeled
& Foerster LLP, Counsel for Netflix, 2 [closing letter] (2010); commercial use. Within artificial intelligence (AI), examples.
http://1.usa.gov/1GCFyXR. machine learning has emerged as the method A diverse array of machine-learning algorithms
15. In re Facebook, Complaint, FTC File No. 092 3184 (2012).
16. FTC Staff Report, Mobile Privacy Disclosures: Building Trust Through
of choice for developing practical software for has been developed to cover the wide variety of
Transparency (FTC, Washington, DC, 2013); http://1.usa.gov/1eNz8zr. computer vision, speech recognition, natural lan- data and problem types exhibited across differ-
17. FTC, Protecting Consumer Privacy in an Era of Rapid Change: guage processing, robot control, and other ap- ent machine-learning problems (1, 2). Conceptual-
Recommendations for Businesses and Policymakers (FTC, plications. Many developers of AI systems now ly, machine-learning algorithms can be viewed as
Washington, DC, 2012).
18. Directive 95/46/ec of the European Parliament and of The
recognize that, for many applications, it can be searching through a large space of candidate
Council of Europe, 24 October 1995. far easier to train a system by showing it exam- programs, guided by training experience, to find
19. L. Sweeney, Online ads roll the dice [blog]; http://1.usa.gov/ ples of desired input-output behavior than to a program that optimizes the performance metric.
1KgEcYg. program it manually by anticipating the desired Machine-learning algorithms vary greatly, in part
20. FTC, “Big data: A tool for inclusion or exclusion?” (workshop,
FTC, Washington, DC, 2014); http://1.usa.gov/1SR65cv
response for all possible inputs. The effect of ma- by the way in which they represent candidate
21. FTC, Data Brokers: A Call for Transparency and Accountability chine learning has also been felt broadly across programs (e.g., decision trees, mathematical func-
(FTC, Washington, DC, 2014); http://1.usa.gov/1GCFoj5. computer science and across a range of indus- tions, and general programming languages) and in
22. J. Podesta, “Big data and privacy: 1 year out” [blog]; http://bit. tries concerned with data-intensive issues, such part by the way in which they search through this
ly/WHsePrivacy.
as consumer services, the diagnosis of faults in space of programs (e.g., optimization algorithms
23. White House Council of Economic Advisers, Big Data and
Differential Pricing (White House, Washington, DC, 2015). complex systems, and the control of logistics with well-understood convergence guarantees
24. Executive Office of the President, Big Data and Differential Processing chains. There has been a similarly broad range of and evolutionary search methods that evaluate
(White House, Washington, DC, 2015); http://1.usa.gov/1eNy7qR. effects across empirical sciences, from biology to successive generations of randomly mutated pro-
25. Executive Office of the President, Big Data: Seizing cosmology to social science, as machine-learning grams). Here, we focus on approaches that have
Opportunities, Preserving Values (White House, Washington,
DC, 2014); http://1.usa.gov/1TSOhiG.
methods have been developed to analyze high- been particularly successful to date.
26. President’s Council of Advisors on Science and Technology throughput experimental data in novel ways. See Many algorithms focus on function approxi-
(PCAST), Big Data and Privacy: A Technological Perspective Fig. 1 for a depiction of some recent areas of ap- mation problems, where the task is embodied
(White House, Washington, DC, 2014); http://1.usa.gov/1C5ewNv. plication of machine learning. in a function (e.g., given an input transaction, out-
27. European Commission, Proposal for a Regulation of the European
Parliament and of the Council on the Protection of Individuals
A learning problem can be defined as the put a “fraud” or “not fraud” label), and the learn-
with regard to the processing of personal data and on the free problem of improving some measure of perform- ing problem is to improve the accuracy of that
movement of such data (General Data Protection Regulation), function, with experience consisting of a sample
COM(2012) 11 final (2012); http://bit.ly/1Lu5POv. 1
Department of Electrical Engineering and Computer of known input-output pairs of the function. In
28. M. Schrems v. Facebook Ireland Limited, §J. Unlawful data Sciences, Department of Statistics, University of California,
transmission to the U.S.A. (“PRISM”), ¶166 and 167 (2013);
some cases, the function is represented explicit-
Berkeley, CA, USA. 2Machine Learning Department, Carnegie
www.europe-v-facebook.org/sk/sk_en.pdf. Mellon University, Pittsburgh, PA, USA.
ly as a parameterized functional form; in other
*Corresponding author. E-mail: [email protected] (M.I.J.); cases, the function is implicit and obtained via a
10.1126/science.aac4520 [email protected] (T.M.M.) search process, a factorization, an optimization

SCIENCE sciencemag.org 17 JULY 2015 • VOL 349 ISSUE 6245 255

 A R TI FI C IA L IN T EL LI GE N CE

procedure, or a simulation-based procedure. Even

when implicit, the function generally depends
on parameters or other tunable degrees of free-
dom, and training corresponds to finding values
for these parameters that optimize the perform-
ance metric.
Whatever the learning algorithm, a key scien-
tific and practical goal is to theoretically character-
ize the capabilities of specific learning algorithms
and the inherent difficulty of any given learning
problem: How accurately can the algorithm learn
from a particular type and volume of training
data? How robust is the algorithm to errors in its
modeling assumptions or to errors in the train-
ing data? Given a learning problem with a given
volume of training data, is it possible to design a
successful algorithm or is this learning problem
fundamentally intractable? Such theoretical char-
Characters Syntax acterizations of machine-learning algorithms and
Visual Motion problems typically make use of the familiar frame-
Dialog Semantics works of statistical decision theory and compu-
tational complexity theory. In fact, attempts to
characterize machine-learning algorithms the-
oretically have led to blends of statistical and
computational theory in which the goal is to simul-
taneously characterize the sample complexity
(how much data are required to learn accurately)
and the computational complexity (how much
computation is required) and to specify how these
depend on features of the learning algorithm such
as the representation it uses for what it learns
(3–6). A specific form of computational analysis
that has proved particularly useful in recent
person 0.92 milk can 1.00 years has been that of optimization theory,
antelope 0.68 milk can 1.00
with upper and lower bounds on rates of con-
vergence of optimization procedures merging
well with the formulation of machine-learning
problems as the optimization of a performance
metric (7, 8).
sunglasses 0.52
As a field of study, machine learning sits at the
crossroads of computer science, statistics and a
variety of other disciplines concerned with auto-
matic improvement over time, and inference and
decision-making under uncertainty. Related dis-
ciplines include the psychological study of human
learning, the study of evolution, adaptive control
theory, the study of educational practices, neuro-
orange 0.73 bird 0.69 science, organizational behavior, and economics.
isopod 0.55 bird 0.95 Although the past decade has seen increased cross-
talk with these other fields, we are just beginning
bird 0.78 to tap the potential synergies and the diversity
of formalisms and experimental methods used
lemon 0.86 starfish 0.67 across these multiple fields for studying systems
that improve with experience.

Drivers of machine-learning progress

The past decade has seen rapid growth in the
ability of networked and mobile computing sys-
tems to gather and transport vast amounts of
data, a phenomenon often referred to as “Big
CREDIT: ISTOCK/AKINBOSTANCI

Data.” The scientists and engineers who collect

Fig. 1. Applications of machine learning. Machine learning is having a substantial effect on many such data have often turned to machine learn-
areas of technology and science; examples of recent applied success stories include robotics and ing for solutions to the problem of obtaining
autonomous vehicle control (top left), speech processing and natural language processing (top useful insights, predictions, and decisions from
right), neuroscience research (middle), and applications in computer vision (bottom). [The middle such data sets. Indeed, the sheer size of the data
panel is adapted from (29). The images in the bottom panel are from the ImageNet database; object makes it essential to develop scalable proce-
recognition annotation is by R. Girshick.] dures that blend computational and statistical

256 17 JULY 2015 • VOL 349 ISSUE 6245 sciencemag.org SCIENCE

considerations, but the issue is more than the work in machine-learning algorithms, theory, and decision trees, decision forests, logistic regres-
mere size of modern data sets; it is the granular, practice. sion, support vector machines, neural networks,
personalized nature of much of these data. Mo- kernel machines, and Bayesian classifiers (1). A
bile devices and embedded computing permit Core methods and recent progress variety of learning algorithms has been proposed
large amounts of data to be gathered about in- The most widely used machine-learning methods to estimate these different types of mappings, and
dividual humans, and machine-learning algo- are supervised learning methods (1). Supervised there are also generic procedures such as boost-
rithms can learn from these data to customize learning systems, including spam classifiers of ing and multiple kernel learning that combine
their services to the needs and circumstances e-mail, face recognizers over images, and med- the outputs of multiple learning algorithms.
of each individual. Moreover, these personalized ical diagnosis systems for patients, all exemplify Procedures for learning f from data often make
services can be connected, so that an overall ser- the function approximation problem discussed use of ideas from optimization theory or numer-
vice emerges that takes advantage of the wealth earlier, where the training data take the form of ical analysis, with the specific form of machine-
and diversity of data from many individuals a collection of (x, y) pairs and the goal is to learning problems (e.g., that the objective function
while still customizing to the needs and circum- produce a prediction y* in response to a query or function to be integrated is often the sum over
stances of each. Instances of this trend toward x*. The inputs x may be classical vectors or they a large number of terms) driving innovations. This
capturing and mining large quantities of data to may be more complex objects such as documents, diversity of learning architectures and algorithms
improve services and productivity can be found images, DNA sequences, or graphs. Similarly, reflects the diverse needs of applications, with
across many fields of commerce, science, and many different kinds of output y have been studied. different architectures capturing different kinds
government. Historical medical records are used Much progress has been made by focusing on of mathematical structures, offering different lev-
to discover which patients will respond best the simple binary classification problem in which els of amenability to post-hoc visualization and
to which treatments; historical traffic data are y takes on one of two values (for example, “spam” explanation, and providing varying trade-offs
used to improve traffic control and reduce con- or “not spam”), but there has also been abun- between computational complexity, the amount
gestion; historical crime data are used to help dant research on problems such as multiclass of data, and performance.
allocate local police to specific locations at spe- classification (where y takes on one of K labels), One high-impact area of progress in supervised
cific times; and large experimental data sets are multilabel classification (where y is labeled simul- learning in recent years involves deep networks,
captured and curated to accelerate progress in taneously by several of the K labels), ranking which are multilayer networks of threshold units,
biology, astronomy, neuroscience, and other data- problems (where y provides a partial order on each of which computes some simple param-
intensive empirical sciences. We appear to be at some set), and general structured prediction eterized function of its inputs (9, 10). Deep learning
the beginning of a decades-long trend toward in- problems (where y is a combinatorial object such systems make use of gradient-based optimiza-
creasingly data-intensive, evidence-based decision- as a graph, whose components may be required tion algorithms to adjust parameters throughout
making across many aspects of science, commerce, to satisfy some set of constraints). An example such a multilayered network based on errors at
and government. of the latter problem is part-of-speech tagging, its output. Exploiting modern parallel comput-
With the increasing prominence of large-scale where the goal is to simultaneously label every ing architectures, such as graphics processing
data in all areas of human endeavor has come a word in an input sentence x as being a noun, units originally developed for video gaming, it
wave of new demands on the underlying machine- verb, or some other part of speech. Supervised has been possible to build deep learning sys-
learning algorithms. For example, huge data sets learning also includes cases in which y has real- tems that contain billions of parameters and
require computationally tractable algorithms, high- valued components or a mixture of discrete and that can be trained on the very large collections
ly personal data raise the need for algorithms real-valued components. of images, videos, and speech samples available
that minimize privacy effects, and the availabil- Supervised learning systems generally form on the Internet. Such large-scale deep learning
ity of huge quantities of unlabeled data raises their predictions via a learned mapping f(x), systems have had a major effect in recent years
the challenge of designing learning algorithms which produces an output y for each input x (or in computer vision (11) and speech recognition
to take advantage of it. The next sections survey a probability distribution over y given x). Many (12), where they have yielded major improve-
some of the effects of these demands on recent different forms of mapping f exist, including ments in performance over previous approaches

Input image Convolutional feature extraction RNN with attention over image Word by word
generation
A
bird
flying
LSTM over
a body
of
14 x 14 feature map
water

A bird flying over a body of water .

CREDIT: ISTOCK/CORR

Fig. 2. Automatic generation of text captions for images with deep networks. A convolutional neural network is trained to interpret images, and its
output is then used by a recurrent neural network trained to generate a text caption (top). The sequence at the bottom shows the word-by-word focus of
the network on different parts of input image while it generates the caption word-by-word. [Adapted with permission from (30)]

SCIENCE sciencemag.org 17 JULY 2015 • VOL 349 ISSUE 6245 257

 A R TI FI C IA L IN T EL LI GE N CE

Topics Documents
gene 0.04
dna 0.02 Topic proportions
genetic 0.01 and assignments
.,,
genes organism genes
survive?

life 0.02 life. genetic

evolve 0.01 computer numbers
organism 0.01 genomes
.,, organisms
sequenced genome
genes computational

brain 0.04
neuron 0.02
nerve 0.01
.,,
predictions

data 0.02 Computer analysis

number 0.02
computer 0.01
.,,

Fig. 3. Topic models. Topic modeling is a methodology for analyzing documents, where a document is viewed as a collection of words, and the words in
the document are viewed as being generated by an underlying set of topics (denoted by the colors in the figure). Topics are probability distributions
across words (leftmost column), and each document is characterized by a probability distribution across topics (histogram). These distributions are
inferred based on the analysis of a collection of documents and can be viewed to classify, index, and summarize the content of documents. [From (31).
Copyright 2012, Association for Computing Machinery, Inc. Reprinted with permission]

(see Fig. 2). Deep network methods are being rithms are developed to optimize the criterion. egy is trained to chose actions for any given state,
actively pursued in a variety of additional appli- As another example, clustering is the problem with the objective of maximizing its expected re-
cations from natural language translation to of finding a partition of the observed data (and ward over time. The ties to research in control
collaborative filtering. a rule for predicting future data) in the absence theory and operations research have increased
The internal layers of deep networks can be of explicit labels indicating a desired partition. over the years, with formulations such as Markov
viewed as providing learned representations of A wide range of clustering procedures has been decision processes and partially observed Mar-
the input data. While much of the practical suc- developed, all based on specific assumptions kov decision processes providing points of con-
cess in deep learning has come from supervised regarding the nature of a “cluster.” In both clus- tact (15, 16). Reinforcement-learning algorithms
learning methods for discovering such repre- tering and dimension reduction, the concern generally make use of ideas that are familiar
sentations, efforts have also been made to devel- with computational complexity is paramount, from the control-theory literature, such as policy
op deep learning algorithms that discover useful given that the goal is to exploit the particularly iteration, value iteration, rollouts, and variance
representations of the input without the need for large data sets that are available if one dis- reduction, with innovations arising to address
labeled training data (13). The general problem is penses with supervised labels. the specific needs of machine learning (e.g., large-
referred to as unsupervised learning, a second A third major machine-learning paradigm is scale problems, few assumptions about the un-
paradigm in machine-learning research (2). reinforcement learning (14, 15). Here, the infor- known dynamical environment, and the use of
Broadly, unsupervised learning generally in- mation available in the training data is inter- supervised learning architectures to represent
volves the analysis of unlabeled data under as- mediate between supervised and unsupervised policies). It is also worth noting the strong ties
sumptions about structural properties of the learning. Instead of training examples that in- between reinforcement learning and many dec-
data (e.g., algebraic, combinatorial, or probabi- dicate the correct output for a given input, the ades of work on learning in psychology and
listic). For example, one can assume that data training data in reinforcement learning are as- neuroscience, one notable example being the
lie on a low-dimensional manifold and aim to sumed to provide only an indication as to whether use of reinforcement learning algorithms to pre-
identify that manifold explicitly from data. Di- an action is correct or not; if an action is incor- dict the response of dopaminergic neurons in
mension reduction methods—including prin- rect, there remains the problem of finding the monkeys learning to associate a stimulus light
cipal components analysis, manifold learning, correct action. More generally, in the setting of with subsequent sugar reward (17).
factor analysis, random projections, and autoen- sequences of inputs, it is assumed that reward Although these three learning paradigms help
coders (1, 2)—make different specific assump- signals refer to the entire sequence; the assign- to organize ideas, much current research involves
tions regarding the underlying manifold (e.g., ment of credit or blame to individual actions in the blends across these categories. For example, semi-
that it is a linear subspace, a smooth nonlinear sequence is not directly provided. Indeed, although supervised learning makes use of unlabeled data
manifold, or a collection of submanifolds). An- simplified versions of reinforcement learning to augment labeled data in a supervised learning
CREDIT: ISTOCK/AKINBOSTANCI

other example of dimension reduction is the known as bandit problems are studied, where it context, and discriminative training blends ar-
topic modeling framework depicted in Fig. 3. is assumed that rewards are provided after each chitectures developed for unsupervised learning
A criterion function is defined that embodies action, reinforcement learning problems typically with optimization formulations that make use
these assumptions—often making use of general involve a general control-theoretic setting in of labels. Model selection is the broad activity of
statistical principles such as maximum like- which the learning task is to learn a control strat- using training data not only to fit a model but
lihood, the method of moments, or Bayesian egy (a “policy”) for an agent acting in an unknown also to select from a family of models, and the
integration—and optimization or sampling algo- dynamical environment, where that learned strat- fact that training data do not directly indicate

258 17 JULY 2015 • VOL 349 ISSUE 6245 sciencemag.org SCIENCE

which model to use leads to the use of algo- vacy. Privacy can be formalized via the notion of reveal most of their genome in the context of
rithms developed for bandit problems and to “differential privacy,” which defines a probabi- research on a disease that runs in their family
Bayesian optimization procedures. Active learn- listic channel between the data and the outside but may ask for more stringent protection if in-
ing arises when the learner is allowed to choose world such that an observer of the output of the formation about their genome is being used to
data points and query the trainer to request tar- channel cannot infer reliably whether particular set insurance rates.
geted information, such as the label of an other- individuals have supplied data or not (18). Clas- Communication is another resource that needs
wise unlabeled example. Causal modeling is the sical applications of differential privacy have to be managed within the overall context of a
effort to go beyond simply discovering predictive involved insuring that queries (e.g., “what is the distributed learning system. For example, data
relations among variables, to distinguish which maximum balance across a set of accounts?”) to may be distributed across distinct physical loca-
variables causally influence others (e.g., a high a privatized database return an answer that is tions because their size does not allow them to
white-blood-cell count can predict the existence close to that returned on the nonprivate data. be aggregated at a single site or because of ad-
of an infection, but it is the infection that causes Recent research has brought differential privacy ministrative boundaries. In such a setting, we may
the high white-cell count). Many issues influence into contact with machine learning, where que- wish to impose a bit-rate communication con-
the design of learning algorithms across all of ries involve predictions or other inferential asser- straint on the machine-learning algorithm. Solving
these paradigms, including whether data are tions (e.g., “given the data I've seen so far, what is the design problem under such a constraint will
available in batches or arrive sequentially over the probability that a new transaction is fraud- generally show how the performance of the learn-
time, how data have been sampled, require- ulent?”) (19, 20). Placing the overall design of a ing system degrades under decrease in commu-
ments that learned models be interpretable by privacy-enhancing machine-learning system nication bandwidth, but it can also reveal how
users, and robustness issues that arise when within a decision-theoretic framework provides the performance improves as the number of dis-
data do not fit prior modeling assumptions. users with a tuning knob whereby they can choose tributed sites (e.g., machines or processors) in-
a desired level of privacy that takes into account creases, trading off these quantities against the
Emerging trends the kinds of questions that will be asked of the amount of data (21, 22). Much as in classical in-
The field of machine learning is sufficiently young data and their own personal utility for the an- formation theory, this line of research aims at
that it is still rapidly expanding, often by invent- swers. For example, a person may be willing to fundamental lower bounds on achievable per-
ing new formalizations of machine-learning formance and specific algorithms that achieve
problems driven by practical applications. (An those lower bounds.
example is the development of recommendation A major goal of this general line of research is
systems, as described in Fig. 4.) One major trend to bring the kinds of statistical resources studied
driving this expansion is a growing concern with in machine learning (e.g., number of data points,
the environment in which a machine-learning dimension of a parameter, and complexity of a
algorithm operates. The word “environment” hypothesis class) into contact with the classical
here refers in part to the computing architecture; computational resources of time and space. Such
whereas a classical machine-learning system in- a bridge is present in the “probably approximately
volved a single program running on a single ma- correct” (PAC) learning framework, which studies
chine, it is now common for machine-learning the effect of adding a polynomial-time compu-
systems to be deployed in architectures that in- tation constraint on this relationship among error
clude many thousands or ten of thousands of rates, training data size, and other parameters of
processors, such that communication constraints the learning algorithm (3). Recent advances in
and issues of parallelism and distributed pro- this line of research include various lower bounds
cessing take center stage. Indeed, as depicted that establish fundamental gaps in performance
in Fig. 5, machine-learning systems are increas- achievable in certain machine-learning prob-
ingly taking the form of complex collections of lems (e.g., sparse regression and sparse princi-
software that run on large-scale parallel and dis- pal components analysis) via polynomial-time
tributed computing platforms and provide a range and exponential-time algorithms (23). The core
of algorithms and services to data analysts. of the problem, however, involves time-data trade-
The word “environment” also refers to the offs that are far from the polynomial/exponential
source of the data, which ranges from a set of boundary. The large data sets that are increas-
people who may have privacy or ownership con- ingly the norm require algorithms whose time
cerns, to the analyst or decision-maker who may and space requirements are linear or sublinear in
have certain requirements on a machine-learning the problem size (number of data points or num-
system (for example, that its output be visual- ber of dimensions). Recent research focuses on
izable), and to the social, legal, or political frame- methods such as subsampling, random projec-
work surrounding the deployment of a system. tions, and algorithm weakening to achieve scal-
The environment also may include other machine- ability while retaining statistical control (24, 25).
learning systems or other agents, and the overall The ultimate goal is to be able to supply time
collection of systems may be cooperative or ad- and space budgets to machine-learning systems
versarial. Broadly speaking, environments pro- Fig. 4. Recommendation systems. A recommen- in addition to accuracy requirements, with the
vide various resources to a learning algorithm dation system is a machine-learning system that is system finding an operating point that allows
and place constraints on those resources. Increas- based on data that indicate links between a set such requirements to be realized.
ingly, machine-learning researchers are formalizing of a users (e.g., people) and a set of items (e.g.,
these relationships, aiming to design algorithms products). A link between a user and a product Opportunities and challenges
that are provably effective in various environ- means that the user has indicated an interest in Despite its practical and commercial successes,
CREDIT: ISTOCK/CORR

ments and explicitly allow users to express and the product in some fashion (perhaps by purchas- machine learning remains a young field with
control trade-offs among resources. ing that item in the past). The machine-learning prob- many underexplored research opportunities.
As an example of resource constraints, let us lem is to suggest other items to a given user that he Some of these opportunities can be seen by con-
suppose that the data are provided by a set of or she may also be interested in, based on the data trasting current machine-learning approaches
individuals who wish to retain a degree of pri- across all users. to the types of learning we observe in naturally

SCIENCE sciencemag.org 17 JULY 2015 • VOL 349 ISSUE 6245 259

 A R TI FI C IA L IN T EL LI GE N CE

In-house apps society to benefit from them, if it wishes to do so.

Cancer genomics, energy debugging, smart buildings
In fact, much of these data are privately held and
Sample clean G-OLA MLBase owned, even though they are data about each of
streaming

GraphX
SparkR
us. Considerations such as these suggest that

Splash
Access and
Spark

BlinkDB interfaces machine learning is likely to be one of the most

Velox
MLPiplines
transformative technologies of the 21st century.
SparkSQL MLIib Processing Although it is impossible to predict the future, it
engine appears essential that society begin now to con-
Spark Core
sider how to maximize its benefits.
Succinct
Tachyon HDFS, S3, Ceph, … Storage REFERENCES
1. T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical
Learning: Data Mining, Inference, and Prediction (Springer,
Resource New York, 2011).
Mesos Hadoop Yarn virtualization 2. K. Murphy, Machine Learning: A Probabilistic Perspective
(MIT Press, Cambridge, MA, 2012).
3. L. Valiant, Commun. ACM 27, 1134–1142 (1984).
AMPLab developed Spark Community 3rd party
4. V. Chandrasekaran, M. I. Jordan, Proc. Natl. Acad. Sci. U.S.A.
110, E1181–E1190 (2013).
Fig. 5. Data analytics stack. Scalable machine-learning systems are layered architectures that are 5. S. Decatur, O. Goldreich, D. Ron, SIAM J. Comput. 29, 854–879
built on parallel and distributed computing platforms. The architecture depicted here—an open- (2000).
6. S. Shalev-Shwartz, O. Shamir, E. Tromer, Using more data to
source data analysis stack developed in the Algorithms, Machines and People (AMP) Laboratory at
speed up training time, Proceedings of the Fifteenth Conference
the University of California, Berkeley—includes layers that interface to underlying operating systems; on Artificial Intelligence and Statistics, Canary Islands, Spain, 21
layers that provide distributed storage, data management, and processing; and layers that provide to 23 April, 2012.
core machine-learning competencies such as streaming, subsampling, pipelines, graph processing, 7. S. Boyd, N. Parikh, E. Chu, B. Peleato, J. Eckstein, in
Foundations and Trends in Machine Learning 3 (Now
and model serving. Publishers, Boston, 2011), pp. 1–122.
8. S. Sra, S. Nowozin, S. Wright, Optimization for Machine
occurring systems such as humans and other organizations, and biological evolution and see Learning (MIT Press, Cambridge, MA, 2011).
animals, organizations, economies, and biological machine learning benefit from ongoing studies 9. J. Schmidhuber, Neural Netw. 61, 85–117 (2015).
evolution. For example, whereas most machine- of these other types of learning systems. 10. Y. Bengio, in Foundations and Trends in Machine Learning 2
learning algorithms are targeted to learn one As with any powerful technology, machine (Now Publishers, Boston, 2009), pp. 1–127.
11. A. Krizhevsky, I. Sutskever, G. Hinton, Adv. Neural Inf. Process.
specific function or data model from one single learning raises questions about which of its po- Syst. 25, 1097–1105 (2015).
data source, humans clearly learn many differ- tential uses society should encourage and dis- 12. G. Hinton et al., IEEE Signal Process. Mag. 29, 82–97
ent skills and types of knowledge, from years courage. The push in recent years to collect new (2012).
of diverse training experience, supervised and kinds of personal data, motivated by its eco- 13. G. E. Hinton, R. R. Salakhutdinov, Science 313, 504–507
(2006).
unsupervised, in a simple-to-more-difficult se- nomic value, leads to obvious privacy issues, as 14. V. Mnih et al., Nature 518, 529–533 (2015).
quence (e.g., learning to crawl, then walk, then mentioned above. The increasing value of data 15. R. S. Sutton, A. G. Barto, Reinforcement Learning:
run). This has led some researchers to begin also raises a second ethical issue: Who will have An Introduction (MIT Press, Cambridge, MA, 1998).
exploring the question of how to construct com- access to, and ownership of, online data, and who 16. E. Yaylali, J. S. Ivy, Partially observable MDPs (POMDPs):
Introduction and examples. Encyclopedia of Operations Research
puter lifelong or never-ending learners that op- will reap its benefits? Currently, much data are and Management Science (John Wiley, New York, 2011).
erate nonstop for years, learning thousands of collected by corporations for specific uses leading 17. W. Schultz, P. Dayan, P. R. Montague, Science 275, 1593–1599
interrelated skills or functions within an over- to improved profits, with little or no motive for (1997).
all architecture that allows the system to im- data sharing. However, the potential benefits that 18. C. Dwork, F. McSherry, K. Nissim, A. Smith, in Proceedings of
the Third Theory of Cryptography Conference, New York, 4 to 7
prove its ability to learn one skill based on society could realize, even from existing online March 2006, pp. 265–284.
having learned another (26–28). Another aspect data, would be considerable if those data were to 19. A. Blum, K. Ligett, A. Roth, J. ACM 20, (2013).
of the analogy to natural learning systems sug- be made available for public good. 20. J. Duchi, M. I. Jordan, J. Wainwright, J. ACM 61, 1–57
gests the idea of team-based, mixed-initiative To illustrate, consider one simple example (2014).
21. M.-F. Balcan, A. Blum, S. Fine, Y. Mansour, Distributed learning,
learning. For example, whereas current machine- of how society could benefit from data that is communication complexity and privacy. Proceedings of the
learning systems typically operate in isolation already online today by using this data to de- 29th Conference on Computational Learning Theory, Edinburgh,
to analyze the given data, people often work crease the risk of global pandemic spread from UK, 26 June to 1 July 2012.
in teams to collect and analyze data (e.g., biol- infectious diseases. By combining location data 22. Y. Zhang, J. Duchi, M. Jordan, M. Wainwright, in Advances in
Neural Information Processing Systems 26, L. Bottou,
ogists have worked as teams to collect and an- from online sources (e.g., location data from cell C. Burges, Z. Ghahramani, M. Welling, Eds. (Curran Associates,
alyze genomic data, bringing together diverse phones, from credit-card transactions at retail Red Hook, NY, 2014), pp. 1–23.
experiments and perspectives to make progress outlets, and from security cameras in public places 23. Q. Berthet, P. Rigollet, Ann. Stat. 41, 1780–1815 (2013).
on this difficult problem). New machine-learning and private buildings) with online medical data 24. A. Kleiner, A. Talwalkar, P. Sarkar, M. I. Jordan, J. R. Stat. Soc.,
B 76, 795–816 (2014).
methods capable of working collaboratively with (e.g., emergency room admissions), it would be 25. M. Mahoney, Found. Trends Machine Learn. 3, 123–224
humans to jointly analyze complex data sets feasible today to implement a simple system to (2011).
might bring together the abilities of machines telephone individuals immediately if a person 26. T. Mitchell et al., Proceedings of the Twenty-Ninth Conference
to tease out subtle statistical regularities from they were in close contact with yesterday was just on Artificial Intelligence (AAAI-15), 25 to 30 January 2015,
Austin, TX.
massive data sets with the abilities of humans to admitted to the emergency room with an infec- 27. M. Taylor, P. Stone, J. Mach. Learn. Res. 10, 1633–1685
draw on diverse background knowledge to gen- tious disease, alerting them to the symptoms they (2009).
erate plausible explanations and suggest new should watch for and precautions they should 28. S. Thrun, L. Pratt, Learning To Learn (Kluwer Academic Press,
CREDIT: ISTOCK/AKINBOSTANCI

hypotheses. Many theoretical results in machine take. Here, there is clearly a tension and trade-off Boston, 1998).
29. L. Wehbe et al., PLOS ONE 9, e112575 (2014).
learning apply to all learning systems, whether between personal privacy and public health, and 30. K. Xu et al., Proceedings of the 32nd International Conference
they are computer algorithms, animals, organ- society at large needs to make the decision on on Machine Learning, vol. 37, Lille, France,
izations, or natural evolution. As the field pro- how to make this trade-off. The larger point of 6 to 11 July 2015, pp. 2048–2057.
gresses, we may see machine-learning theory this example, however, is that, although the data 31. D. Blei, Commun. ACM 55, 77–84 (2012).
and algorithms increasingly providing models are already online, we do not currently have the
for understanding learning in neural systems, laws, customs, culture, or mechanisms to enable 10.1126/science.aaa8415

260 17 JULY 2015 • VOL 349 ISSUE 6245 sciencemag.org SCIENCE