Science of Learning: History

brewerobstructionAI and Robotics

Nov 7, 2013 (3 years and 7 months ago)

55 views

Science of Learning: History

A workshop held at NSF, 4
-
5 October 2012

Submitted by the Steering Committee: David Lightfoot
(PI), Ralph Etienne
-
Cummings,
Morton Gernsbacher,
Eric Hamilton, Barbara Landau, Elissa Newport and
David Poeppel



































2


Science of Learning Workshop: History


1. Introduction

For a long time NSF has

supported work on learning,

through regular programs
in SBE, CISE and EHR and
also
through special large initiatives like
the
L
earning
and Intelligent
S
ystems component of the two
-
year Foundation
-
wide program in

Knowledge
and Distributed Intelligence
(KDI)
in the late 1990’s.

In its support
,

NSF
has
responded to ground
-
breaking shifts in our understanding of learn
ing
as thinking moved beyond B. F.
Skinner’s

long
-
dominant behaviorist para
digm
,
from
a single associationist approach toward an appreciation of the complexity


and potential multiplicity


of learning mechanisms
.
As one

example, t
he 1973
Nobel Prize in Physiology (to Konrad Lorenz, Niko

Tinbergen and
K
arl

von
Frisch)
marked the discovery and description of what were originally called
“innate releasing mechanisms” in ethology: an external triggering stimulus
releases a developmental program that allows the organism to learn highly
specifi
c actions or representations; it requires a well
-
articulated, genetically
specified scaffold that is triggered by input.
Appendix 1 provides a brief
bibliography
of
some of
the influential ground
-
breaking work

that

brought about
the

shifts

in our understa
nding of learning processes
.


NSF has construed learning broadly, dealing with the cognitive and neural basis
of human lear
ning, learning in other animals

and computer models of learning.
In 2003
it

established the Science of Learning Centers (SLC)
progra
m
. The goal
was to stimulate and integrate research in the science of learning, dealing with
the cognitive and neural bases of learning (as distinct from the more education
-
driven “learning sciences”); to connect the research to scienti
fic, technological,

educational

and workforce challenges; and to enable research communities to
capitalize on new opportunities and discoveries
.

The thinking was that the
complexity of the
se

goals required expertise from various disciplines and
int
egrative research agendas
that we
re beyond the capabilities of individual
investigators or small groups. The longer durations of funding and the stable
environments of centers would provide incentives for committed, long
-
term
interactions among researchers to reconceptualize their

thinking beyond the
paradigms of traditional disciplines.
The first solici
tation
is at
SLC Solicitation

and
the six
centers are listed in Appendix 2
.


The SLC Program has represented a big investment in the human sciences
broadly and in the multidisciplinary science of learning involving several of
the
NSF directorates. As the centers begin to phase down toward the expiration of
NSF support after ten ye
ars, the time has come to think about the future of the
science of learning and, to this end, two two
-
day workshops are being held,
The
Science of Learning: History and Prospects
.

This report covers the first workshop,
held at NSF on 4
-
5 October 2012 and
dealing

with what has
been achieved over
recent decades in the science of learning
, particularly in the last ten years
. The
s
econd workshop, to be held at NSF on 28 February and 1 March 2013, will

3

consider what we can look fo
rward to over the next ten yea
rs

in terms of
opportunities and threats and will

be a forum to brainstorm mechanisms for
how

work on

the science of learning might be supported and funded over the
coming decade, outlining strategies and objectives.


2. Organization of the workshops

A S
teering Committee of leading figure
s in work on learning is guid
ing

the
organization of

both workshops

and is writing

the reports
:
Ralph Etienne
-
Cummings (Johns Hopkins),
Eric Hamilton (Pepperdine),
Elissa Newport
(Rochester and now Georgetown),
David Poep
pel (NYU and recent member of
the SBE Advisory Committee), and current members of the SBE AC, Morton
Gernsbacher (Wisconsin) and Barbara Landau (Johns Hopkins)
.


For the first workshop, s
ix speakers were invited to address topics in the science
of
learning: Michael Stryker fr
om UC San Francisco on

neural plasticity, Nit
in
Gogtay from NIMH on

abnormal and normal brain development, Ranu Jung from
Florida International
on

motor control
learning linked to rehabilitation, Sharon
Goldwater from Edinburgh

Univer
sity on

computational modeling and large
-
scale data
-
mining, Linda Smith

from Indiana University on cognitive
development, and David Andrews from Johns Hopkins on learning and
education. Soo
-
Siang Lim
from NSF
was invited to discuss infrastructure
d
eveloped by the SLC Program through the six centers and six representat
ives
from the centers were ask
ed to speak about achievements and challenges in
the
focal area of their center:
Nora Newcombe from SILC on spatial learning, Pat
Kuhl from LIFE on social

foundations of

learnin
g, Ken Koedinger from PSLC on
computational

models and robust learning
, Barbara Sh
inn
-
Cunningham from
CELEST on brain
-
inspired technologies
, Gary Cottrell from TDLC on timing
elements in learning,
and Laura
-
Ann Petitto from VL2 on vis
ual learning and
signed languages. All speakers were invited to identify two signal achievements
and two challenges in the areas they were address
ing
.


Presenters all sent in one
-
pagers in advance of the meeting
,

listing their main
points and providing l
inks to publications

(Appendix 5
)
. There was extensive
discussion: ten minutes after each presentation, half an hour at the end of the
first day
,

and then structured discussion for the whole of the morning on the
second day. The
list of participants
is i
n Appendix 3 and the program in
Appendix 4

(with links to the one
-
pagers)
.


3.
The science of learning


The presentations from in
vited speakers and from representative
s of the
existing Science of Learning Centers covered broad territory, raising the
question of what we mean by

“learning”

and what has been discovered about its
processes and mechanisms.
One relatively broad idea, provided by Michael
Stryker's contribution, is
that learning consists of some ‘
reasonably specific set
of changes in neural
connections cor
responding to the thing learned
.


It is
notable that this idea does not constrain learning to changes that depend on

4

experience per se. For example, formation of structure in the developing visual
system occurs as a consequence of both spo
ntaneous neural activity and
exposure to structured patterns of information available to the organism
from
the environment. A slightly narrower idea is that learning encompasses
experience
-
dependent change. Even here, the range of changes that are
conseq
uent upon experience, the kinds of experience that create change, and the
timetable on which these changes can occur constitutes vast territory.
As a
consequence, it is

likely that the mechanisms underlying learning could be quite
varied. Consider, for
example, the infant who
learns

to reach

and grasp objects;
the toddler

who
learns to talk and understand
; the child who
learns
to count or to
read; the adolescent who
learns
to drive; the adult who
learns
to re
-
use his or
her limbs after stroke. Moving in
to the realm of machine learning, consider the
machine that
learns

to translate an unknown language,
learns
to diagnose a
tumor type on the basis of brain images, or
learns
to play Jeopardy, and compete
with human experts.


The vast territory that comprise
s human learning can be organized to some
degree by considering evolutionary foundations, the specific domains of
learning, and likely mechanisms underlying learning in a given domain.
Evolutionary foundations suggest that some aspects of human learning a
re likely
to be continuous with other species (e.g. development of visual
-
motor
coordination, tool use, number, navigation)
,

while others are likely to be distinct
from that of other species (e.g. human language, formal use of symbol systems).
Still other
s will likely be hybrids, in which some foundational aspects of the
system are shared by many species while other accomplishments require formal
tutoring only available to h
umans. Number constitutes a good example:

while
fundamental aspects of numerical s
ensitivity are shared by
other
species, only
humans master algebra

(Dehaene 1997)
.


Domain
-
specific structures vary considerably, suggesting that some domains
may engage distinct learning mechanisms.

For example, navigation in all species
requires that
the organism keep track of its current location as it moves through
space; for many species, this in turn depends at least partly on the mechanism of
dead
-
reckoning, which allows the animal to keep track of its changing
location as
it moves (Gallistel

1990
) and this supports the ability to form a map of the
environment. The distinctly different case of language acquisition has been
subject to intense controversy, with solid evidence now showing that aspects of
the learning problem may depend on quite gener
al statistical learning
mechanisms (e.g. parsing the speech stream, S
affran, Aslin & Newport

1996) but
other aspects of learning in syntax and semantics
are
still unexplained by such
general mechanisms.

It has been discovered (i) that language acquirers can
entertain multiple representations
of a syntactic string and (ii) that the
representations entertained sometimes go against the statistics of the input: that
is, learners entertain highly constrained

options that are only in part dr
iven by
properties of the input.
In addition
,

learning mechanisms may vary depending
on the knowledge domain and
,

therefore, the com
putational problem to be

5

solved.


L
earning mechanisms have also been categorized by scient
ists at a more
macro level, into those that appear to require explicit (conscious) learning (as in
learning a list of new word pairs by reading them out loud) or implicit
(unconscious) learning (as in learning the properties of "outdoor vs. indoor
scenes"
by passively observing many exemplars, and constructing summaries

of
their statistical structure
)
.


These basic organizational cuts are surely inadequate to capture the full richness
of learning. Moreover, they may leave as
ide many kinds of change that
-

a
lthough
they might not be part
of the natural kind "learning"
-

will likely shed light on the
breadth of changes that any science of learning will want to capture. These
include such cases as the changes to the visual system underlying the
development of
binocular vision; changes to the developing brain that occur as
information is recruited, manipulated and stored; changes to memory during the
life
-
span and in the diseased brain;
and
changes that occur during rehabilitation
after brain injury. The vast t
erritory of learning requires not just a single science
of learning, but
,

more likely, multiple
sciences

of learn
ing.


4.

Some history

The last heyday of learning theory was during the 1940’s and 1950’s, when the
study of learning was dominated by ass
ociat
ionist theories

that proposed a few
general principles that would explain all types of learning, across domains and
species. That optimistic view waned with two striking findings. First, the
seminal work of John Garcia showed that, even in rodents and bi
rds, simple
principles of conditioning were invaded by species
-
specific biases and innate
constraints on what could be readily learned. Second, Noam Chomsky
profoundly altered our understanding of cognition, suggesting that there are
abstract universal pr
inciples of human language and arguing that language
learning (and other types of learning) is made possible by constraints on the
types of patterns that can be learned and processed. Together these lines of
work, and others that followed in fields from p
sychology to computer science,
have suggested that learning systems operate successfully by being quick to
acquire certain types of information
-

and correspondingly slow or entirely
fail
ing

to acquire other types.

Surprisingly, for a few decades after t
hese claims appeared, the study of learning
continued within linguistics and computer science but, without a search for
general principles, languished within psychology. Departments of psychology
that had always offered courses on ‘
l
earning’ and had progr
ams of graduate
study focused on a
nimal learning ceased to offer

these specialties. But in more
recent years, several important findings have revitalized interest in the study of
learning, which is now one of the cutting edge fields within cognitive sci
en
ce.
First, while the
Chomsk
y
an analysis

of specialized learning modules

has become
richer and deeper, challenges

have come from the study of neural networks, and
the controversies surrounding this work, both
from supporters and critics, have

helped to put the study of learning back in the center of the cognitive

and neuro
-


6

sciences. Second, discoveries within neuroscience of some of the cellular
-
molecular and systems
-
level underpinnings of learning


from LTP and NMDA
receptors to studies of

the hippocampus and other memory systems


have
begun to shed light on the mechanisms by which experience alters the brain.

Third, the field of infancy has provided remarkable findings of very early human
cognitive capacities and also very early capaciti
es for learning, even including
prenatal learning. Fourth, the field of m
achine learning has undergone
revitalization, providing a wealth of computational models for how human (and
non
-
human) learning might in principle work.

Among the many important disc
overies of recent years are the following:



Developmental and adult plasticity
: We have learned not only that the
brain is particularly plastic and susceptible to environmental influences
early in life, but also that it is still, to some degree, plastic even in
adulthood. Adult
plasticity is reduced as compared with early
developm
ent, but some of the same mechanisms for plastic change are
still present in the adult brain


and new findings show even how to re
-
open critical periods for plastic change in mature organisms.



Cross
-
species comparisons of learning

and the

evolution of lea
rning
mechanisms:


We have also learned that many me
chanisms of learning
are shar
ed across species, and have begun to understand as well the
arenas in which learning differs or has evolved
differently
across
species and domains.

An excellent summary of wo
rk on non
-
human
animals i
s

Gallistel et al.’s 1991 landmark review.

Evidence now
abounds that, for many important learning problems, most species
begin life equipped with structures and mechanisms that guide
learning.


Examples include the barn owl, equipped with a specialized
learning mechanism that calibrates its sound localization circuitry as it
grows; migratory songbirds that are capable of representing the spatial
arrangement of the stars in order to direct their
initial flights; and ducks
that compute the relative distribution of foods so that they can select
the optimal location for foraging.



Mechanisms of learning,
integrating from cells to behavior
:

In several
important systems in animals



particularly in the
sensory and motor
systems


there has been remarkable progress in understanding

both
the effects of experience and the cellular
-
molecular changes that
mediate them in shaping neural circuits.

o

From early seminal work
in

vision, we know two important
princ
iples. The early
findings of
Hubel

&
Wiesel
(
1962
)
show that
early visual input to the two eyes in cats can permanently alter the
size of the neural regions devoted to each eye, and also their relative
dominance in binocular vi
sion (a critical period effe
ct

of input on
neural circuits). We also know that the broader organization of
visual cortex, as well as other sensory and motor cortices,
is

7

fundamentally “
topical
,


with a consistent mapping of
the receptor
surface (e.g.

from left to rig
ht in the visual

system,
from low to high
pitch

in the auditory system
) onto the corresponding layout of the
primary cortical areas.

o

From the work of Knudsen

(
1999,
2004
)
,
Carr &
Konishi

(
1988
)

and
others

on barn owls
, we have learned
how

early auditory experience
can alter sound localization
;

th
e mechanism
s

by which sound
localization is mediated
, through

cleverly evolved
simple neural
circuits

(delay lines);
and the rich ways in which these m
echanisms
can and cannot be altered
throughout

life,

by experience
with flight
and

localization

of prey
.

o

We have learned, from the work of Merzenich

et al.

(
1983
)
,

about
reorganization of somatosensory cort
ex

in primates
that can occur
with experience
using the hands, even in adulthood
.

o

While these ma
tter
s are much more difficult to investigate in
humans


and links between cellular circuitry and behavior are at
present out of reach
-

t
he study of language is one prominent arena
in human cognitive science
in which critical periods and plasticity
early
versus late in life has been the subject of important and
sophisticated investigation.



Types of learning
and

memory
:

As interest in fundamental

principles of
learning has been revived in basic behavioral research, a diversification
of types of learning has b
een explored. Some cognitive

scientists
distinguish procedural and declarative learning, the learning of
procedures (such as how to ride a b
icycle or compute a square root)
versus the learning of information (such as the capital of Brazil or the
color fuchsia).

Other

scientists distinguish between short
-
term learning
(including the maintenance of knowledge in so
-
called working memory)
and lon
ger
-
term learning.

Engle et al. 1999

demonstrated a strong
correlation between the ability to quickly store and accurately retrieve
recently learned information and standardized assessments of fluid
intellig
ence.

Still other

scientists distinguish betwee
n implicit learning,
obtained without conscious awareness or notable effort, and explicit
learning, which requires effortful encoding and rehearsal.



T
here are many different
type
s of learning
.
Domains of knowledge
such
as language, space, number

and
,

likely, social interaction, are well
-
structured, but quite different from each other, and learning in each
domain depends on having some initial structural biases. The biases for
each domain are qualitatively differe
nt

and there is no necessary reason
fo
r the mechanisms underlying our ability to produce and understand
complex sentences to be identical to the mechanism
s

underlying our
ability to navigate through space or decide whether a con
-
specific

8

should be trusted. Of course, the question of whether d
omain
-
general
mechanisms also play an important role in learning in every domain,
and how these mechanisms interface with domain
-
specific mechanisms,
is still very much under debate.


We have
also learned that many case
s of learning are subserved by
anatom
ical regions that support different types of operations. For
example, the functional organization of primary visual cortex (V1)
includes both ocular dominance columns and orientation
-
selective
pinwheels. The same function can also be implemented by diffe
rent
neural mechanisms that support different algorithms
:
consider the
different neural mechanisms for sound localization

in different animals

(Groth
e

2003). Learning also involves strong feedback and feed
-
forward mechanisms, with no clear sequential
processing from simple
to complex function. Areas previously thought to perform very low
-
level processing are also modulated by high
-
level

information (e.g. V1 is
modulated by online language processing and linguistic experience; see
Dikker et al. 2010
)
.





Machine learning and learning in silicon:
Alan Turing’s proposition in
1950 that

the only way to determine if a machine can actually learn is if
we communicate with it and cannot distinguish it from another human

remains

the open challenge to the mach
ine learning community. The
recent IBM Jeopardy
TM

player, “Watson”, made some progress towards
responding to this challenge, however, a major gap still exists between
what humans and other living organisms can learn, and learning
capabilities of machines.


Some landmark developments followed

Turing’s proposition:

o

In 1952 Arthur Samuel (IBM) wrote the first learning
-
based game
-
playing program, for checkers, to achieve sufficient skill to challenge
a world champion. This lead to the ELIZA system in the
early 60’s
,

which simulated a psychotherapist by using tricks like string
substitution and canned responses based on keywords. When the
original ELIZA first appeared, some people actually mistook her for
a
human.

o

In 1957 Frank Rosenblatt invented the Per
ceptron, a simple linear
classifier
, which, w
hen configured into networks, could solve hard
problems. Minsky and
other
s in the 1960s challenged the
effectiveness of the approach by showing that simple problems
could disrupt its functionality (Minksy & Pape
rt 1969). Nonetheless,
researchers continued to improve the efficacy of perceptrons,
leading to today’s Deep Learning Architectures (Hinton et al. 2006)
and Support Vector Machines (Vapnik et al. 1997).

o

Since
the
late 1980s, engineers have been modeling
the nervous
system in silicon using integrated circuits (Hopfield 1987
,

Mead &


9

Mahowald

1988). More recently, learning silicon synapses, using
Spike
-
Timing
-
Dependent Plasticity (STDP), have

also been included
in these so
-
called
neuromorphs

(Indiveri et al
. 2006
)
.
We are now
witnessing the advent of ultra
-
large scale models of the nervous
system in silicon, models that use billions of neurons and trill
ions of
learning synapses
(Ananthanarayanan et al. 2009)
.




Sleep and consolidation:

There is now
overwhelming evidence that
sleep

enhances memory consolidation
-

that is, it strengthens memories
and makes them resistant to disruption. Conversely, sleep deprivation
results in memory deficits. The exact mechanisms underlying these
results are not fully

understood, but theory and empirical evidence
suggest that this is at least partly due to the "replay" of experiences
during the sleep cycle, with the hippocampus playing a significant role.
A review by Ellenbogen, Payne & Stickgold (2006) concludes ‘tha
t sleep
leads to improved performance in memory recall; that sleep renders
memories resistant to subsequent interference; that the resistance to
interference lasts throughout the subsequent waking period; that
certain stages of sleep correlate with perform
ance improvements on
certain tasks; that the hippocampus replays information during sleep;
and that the behavioral improvements correlate with hippocampal re
-
activation. Given this evidence, we believe the most parsimonious
conclusion is that there are spe
cific, sleep
-
dependent, neurobiological
processes that directly lead to the consolidation of declarative
memories.’



Statistical learning, Bayesian hierarchical models, information theoretic
approaches to learning:

Surprising discoveries have been made abo
ut
the abilities of human learners to acquire and use complex probabilistic
information
-

about the organization and sequencing of linguistic
elements, as well as the causal structure of the world (Saffran, Aslin &
Newport 1996, Tenenbaum et al. 2011). Ea
rlier approaches assumed
that human learners could not compute or retain such complex
information about speech corpora or causal scenes; and, in the absence
of such information from the environment, theories would need to
assume more complex innate structu
re or more limited abilities to
learn. However, recent research, some using Bayesian models, has
shown that both infants and adults can implicitly and rapidly compute
rich information about the structure of the environments to which they
are exposed, and
has stimulated the development of old and new types
of information theoretic approaches to learning. Domain
-
general
probabilistic learning mechanisms are critical but they interact with
biological constraints and research focuses on the interplay between
constraints and learning.



Number, language acquisition, and space as domains of development

10

and learning:

Following
the seminal work of Tinbergen (1951) and
Chomsky (1959, 1965), generalist theories of learning and development
were challenged; theoretical

frameworks developed in the second half
of the 20th century stressed that different domains of knowledge
require the construction of qualitatively different representations and
associated learni
ng mechanisms (Gallistel et al.

1991 for review).
Remarkable

empirical discoveries

within number, space, language

and
social interaction illustrate that domain
-
specific foundations of human
knowledge can be observed even quite early in development. For
example:

o

Numerical cognition is now known to be rooted in a sp
ecies
-
general
ability to use the approximate number system,
which permits
infants, children

and adults to estimate
small and
large
numerosities. This system engages specific brain circuits in humans.
Later acquisition of numerical competences involving la
rge exact
number
s is

thought to be unique to humans
. These ideas have
generated an active search

for

mecha
nisms underlying the varieties
of numerical knowledge

(Dehaene 1997)
.


o

Spatial representation, especially navigation, has long been known
to have a
specific neural foundation, with the hippocampus and
related structures playing a crucial role. Computational sub
-
divisions of the navigation problem have revealed a range of
structural components, with the problem of re
-
orientation revealing
both solid co
ntributions available across species and in human
s as
young as infants, as well

as

learned contributions that change
navigation behavior
over development

(
O'Keefe & Nadel

1978
,

Gallistel
1990
,

Hermer & Spelke
1996
,

Newcombe & Ratliff

2007
).

o

The full acquis
ition of language requires the construction of highly
specific knowledge; there is on
-
going debate about the relative roles
of domain
-
general processes (such as statistical learning) and other
learning mechanisms in acquisition.

Problems such as parsing o
f the
speech stream and formation of basic categories engage powerful
statistical learning mechanisms

even in infancy
, whereas problems

such as
the acquisition of
complex syntax and semantics remain
challenges to existing proposed learning mechanisms

and r
equire
postulating biological constraints on what is learnable through
variable environmental input
.

o

The role of social interaction in learning is now known to be a
powerful force in early development, with recognition that the
human ability to imitate a
nd to explicitly teach each other may
separate

humans from other species, resulting in learning systems

that go beyond those of other species

(
Hermann

et al. 2009)
.

5
.

Tools for investigating l
earning

The past two decades have
also
witnessed great
advances in the tools that
scientists can use to explore the mechanisms underlying and promoting human

11

learning.


Foremost has been the advent of non
-
invasive brain imaging.

Prior to
these advances, explorations of the neural bases of cognitive processes
as

ubiquitous as language, memory

or attention, were dependent on inferences
drawn from accidents of nature: persons who had experienced brain damage.

Functional magnetic resonance imaging (fMRI), as well as the increasing
developments of other non
-
invasi
ve brain imaging techniques such as NIRS
(Near
-
Infrared Spectroscopy), allow us to watch the intact brain at work and to
compare function across groups or within an individual across time in children
as young as preschool age.


These methods comple
ment the

use of EEG
(Electroencephalography), which is becoming increasingly more sophisticated in
its applications, and MEG (Magnetoencephalography).


TMS (Transcranial Magnetic S
timulation), which
stimulates

neurons electrically

as a means of eliciti
ng, modifyin
g

and inhibiting responses, and TDCS
(Transcranial Direct Current Stimulation) are also
employed
increasingly as
tools for exploring the neural bases of learning.

P
utatively

non
-
invasive, these
techniques have been heralded as improving lea
rning (of video

games),
decreasing

train
ing time for pilots

and improving

plasticity following brain
injury.


Advances in structural brain imaging have also aided our understanding of the
brain bases of learning.


In addition to identifying attributes of volume and
shape
, tools such as DTI (
diffusion tensor imaging) allow

visualization, in three
dimension
s, of
white matter tracks and fibers.

For example, a recent DTI study
of

elite gymnasts identified the ‘
neuroanatomical adaptations and plastic
changes [that] occur in
gymnasts' brain anatomical networks either in response
to long
-
term intensive gymnastic training or as a
n innate predisposition or both’

(W
ang et al.

2012).


The Internet is also providing tools for investigating learning.

Amassing large
data sets, such a
s those contained in CHILDES (the Child Language
Data
Exchange System), TalkBank

and the Linguistic Data Consortium, is aided by
resear
chers’ ability to upload, store

and download through cloud computing.

Internet
-
based data collection, through platforms
such as Mechanical Turk and
Label Me, is revolutionizing the breadth of research participant samples.

And
Internet sites, such as NeuroSynth.org, which automatically extracts and
syntheses fMRI data from published articles, are allowing researchers to
com
pare studies and findings in a fraction of the time required previously.


6
. I
nterdis
ciplinarity


The construct of learning was once covetously contained wit
hin its own
discipline. In
past decades it has attracted multiple disciplines, often working
together. The field is replete with examples

but we will illustrate it by
considering
a marriage of engineering and cognitive psychology.



12

New cyberinfrastructure has had transformative consequences throughout
scientific research, as is well known. The t
ransformative consequences for work
on learning have taken two quite different forms: providing models of learning
and tools for learning
, notably

search tools.


Machine learning and their silicon implementation counterparts have very
different roots. Mac
hine learning algorithms emerged primarily out of
mathematics, and their goal has been to perform a task, regardless of the
similarity of the methodology to biological learning. On the other hand, silicon
models of learning have been primarily driven by t
he desire to mimic biological
learning in silicon. The two approaches started to converge in the 1980s, when
Sejnowski and Rosenberg developed a parallel compute
r that learned to read
and pronounce

text, and Hopfield presented his associative memory netwo
rk of
resistors and comparat
ors (Sejnowski & Rosenberg 1987,

Hopfield 1987
). In
both cases, deeply mathematical constructs were readily implemented with
digital (former) and analog (latter) electronic circuits, and both systems
purported to have some rele
vance to biological learning. Machine learning, in
the form of perceptron learning, continued to develop mainly along
mathematical and computer science lines (e.g. support vector machines and
deep belief networks), however, the silicon implementation side

of the equation
continued to look to neuroscience for inspiration.


The field
s

converged further with the discovery of spike
-
timing
-
dependent
-
plasticity (STDP) by Markram and colleagues (1997) and Long
-
Term Depression
and Protentiation (LTD & LTP) by Fe
ldman (1999). These results, which
provided a mechanism for strengthening and weakening of biological synapses,
also correlated with the biological learning theory proposed

by psychologist
Donald Hebb

in the 1940s. Furthermore, the elegance of STDP, and
the ease
with which it can be implemented with digital or
analog integrated circuits, has

led to a large number of silicon implementation
s

of learning in a
rtificial neural
systems (Ananthanarayanan et al. 2009
).


While STDP explained how individual conne
ctions between individual neurons
can be modified, statistical learning, e.g. hierarchical
B
ay
e
sian networks,
provided a means for inferring complex relationships between high dimensional
input data streams and output actions at much higher levels than the

individual
neurons (Lee & Mumford 2003). Again this represented a further conversion of
cognitive science and mathematics to attempt to explain how learning occurs in
cortex. What is evident in these developments is that the study of machine
learning an
d/or their implementation in electronic hardwa
re requires the
interaction of myriad

fields:

mathematics, computer science and engineering,
neuroscience, psychology and cognitive science play central roles in developing
theory, models, algorithms and practi
cal implementations.


Machine learning is playing increasingly indispensable roles in our everyday
lives. We routinely use search engines on the web to find information and get

13

recommendations (e.g. Google
, Bing, Netflix). We use voice
recognition/tran
slation software to interface with our computers or to interact
with customer service agents (e.g. Siri, Android, Google translate, Dragon).
Nowadays, machine learning is even used to automate the grading of
standardized exams. Furthermore, such algorithm
s are also becoming more
prominent in healthcare for mining databases, for bioinformatics and imaging
retrieval and processing, and for correlating tests results with pathologies. In
general security systems, they are used to identify credentialed users a
nd
anomalies in their actions that may indicate potential threats to the systems.
Machine learning also plays a significant role in the financial sector, predi
cting
trends, identifying risks

and suggesting investment strategies. With further
development,

improved accuracy and faster execution, they will become even
more integrated in society so that machines will become even more
indistinguishable from other humans, thus meeting the challenge that Turing
posed for the community in 1950.


7
. Education an
d society

We have been describing and dissecting learning as it is situated within an
individual organism, but the process of learning plays a powerful role across
individuals into groups and societies at various levels.
Learning shapes and is
shaped by s
ocial dynamic and some aspects of learning require attention to the
socio
-
cultural context.


Many of our society’s most precious institutions


economic liv
elihood, scientific
achievement

and productive international relations


rely on a learned citizenry
.
One of the most obvious applications of the science of learning is the education
system. However, although the science of learning, as a discipline, draws
scientists from a wide range of fields


cognitive psychology, educational
psychology, computer s
cience, engineering, neuroscience, sociology,
anthropology and education


the interface betw
een educators and scientists has
not yet reached equilibrium. T
he applications of science to formal learning in
school environments
are understood at only
primitive levels, and scientific
findings about connections between learning and usable strategies in education
are continually evolving
.


Nonetheless, part of the mandate of the centers was
to connect the research to
scientific, technological, educationa
l and workforce challenges

and
the centers as
a group have worked

with educators and schools;
Pittsburgh’s LearnLab, in
particular, has
provided a basis for experimental work and other centers have
generated

innovative forces in school systems
. In additio
n, there is a scientific
basis for certain principles

ripe for ap
plication in the schools, including

the
following

principles that have animated education researchers in recent years
and have been summarized

by Sawyer (2008, p58):


-

Acquiring deeper
conceptual understanding is more valuable than
memorizing superficial facts and procedures.


14

-

Learning connected and coherent information is more valuable than
learning information compartmentalized into distinct subjects
.

-

Gaining knowledge that is authe
ntic to its context is more valuable than
executing “decontextualized classroom exercises
.


-

Learning in collaboration is often superior to learning in isolation
.


Sawyer further proposes that learning environments
that

have the following
characteristics
will be most effective:

-

Each learner receives a customized learning experience.

-

Learners can acquire knowledge whenever they need it from a varie
ty of
sources: books, web
sites

and experts around the globe.

-

Students learn together as they work collab
oratively on authentic,
inquiry
-
oriented projects.

-

Tests should evaluate the students’ deeper conceptual understanding,
the extent to which their knowledge is integrated, coherent and
contextualized.


When viewed from this perspective, learning extends
beyond the bounds of a
classroom to community environments such as museums and extra
-
curricular
activities, and the process of education expands beyond the traditional structure
of schools.

Moreover, learning is situated beyond the standard transmission o
f
information from a teacher to a student, expanding the societal contributions to
education.


8
. The role of the centers

The science of learning has been a vibrant field over the last decades, as shown
by the recent conceptual shifts (sections 1 and 3)
and new tools and new
cyberinfrastructure (section 5). We understand more about how learning takes
place and, as a result, we learn more about the very nature of lan
guage, audition,
vision, memory

and other mental faculties by considering how they change
over
time and through experience. Many important discoveries have been made and
section 4 and the first six one
-
pagers pro
vide a sample (Appendix 5
).


In addition, while much has been accomplished in work on learning generally,
the centers themselves have

been productive. Each representative at the
workshop provided two achievements (
see the one
-
pagers in Appendix 5
); one
presentation (Soo
-
Siang Lim) described new infrastructure that the centers have
generated. In their most recent reports, SLC’s have id
entified approximately fifty
important findings, as well as reconceptualizations of commonly held
assumptions in the field, new research communities, new tools, and training
student researchers with fresh patterns of inquiry. A brief synopsis of these

sel
f
-
reported

center accom
plishments appears in Appendix 6
, with links to fuller
reports
.


The synergistic focus of the SLC cooperative agreements tasked each center with
the challenges of collaboration. SLC researchers were pressed to cross
boundaries and

to seek insights that extended beyond the perspective of

15

individual research communities or teams. The boundaries that SLC researchers
have crossed are varied and include, for example, temporal scales (e.g.
microsecond activation patterns to small graine
d to more extended task
performance), spatial scales (e.g. cellular to structural to regional to functional to
extended task performance), and scientific domains (e.g. from neuroscience to
cognitive psychology, to data
-
mining m
ethods, to impairment therapi
es

and
education research more broadly).



While the term “translation” often is associated with a notion of making research
on learning applicable in contexts such as education (i.e. from basic research to
practice), a primary task of the SLCs has been t
o build translation between basic
research communities and broader ranges of activity, including industry and
business, partly to generate new ways of viewing research problems. This has
been based at least in part on an expectation that researchers colla
borating
across boundaries would create different and more advanced landscapes for the
entire field.


Cross
-
boundary endeavors were often expressed thematically by individual SLC
emphases. SILC, for example, has sought to connect the cellular and functio
nal
aspects of the malleability of spatial cognition with scientific learning. In so
doing, it views itself as founding a new research community emphasizing spatial
cognition’s role in STEM education. TDLC examines the relationships between
finely graine
d neural event sequencing and signal systems and higher level
functioning in learning and cognitive therapies. The LIFE Center has made
significant breakthroughs in areas such as early childhood functional brain
imaging and in tracing connections between
social patterns and language
acquisition, while seeking to advance a broader and cross
-
disciplinary
conceptualization of informal human learning. In collaboration with TDLC, the
LIFE Center has charted principles to guide a multi
-
tiered and cross
-
discipli
nary
science (or sciences) of learning. Cross
-
boundary research programs especially
characterize CELEST, the SLC most heavily focused on cognitive neuroscience.
Its research on memory structures, for example, has made significant strides in
connecting br
ain activation patterns to the cognitive functions that organize and
process visual and auditory stimuli. These research threads are part of a fuller
body of inquiry in human and animal learning that explores connections
between regions or functions of th
e brain. LearnLab at PSLC has approached
integrative research in part by building what is the world’s largest repository of
cross
-
boundary research related to
in situ

classroom learning. LearnLab’s own
r
esearch studies are consistently

marked by analysis

of fine
-
grained data in
service of developing larger grained models and theories that seek to test causal
mechanisms of human learning processes and related socio
-
affective dynamics.


While collaborative multidisciplinary research is given privileged stan
ding in the
SLC network, each SLC has also carried out significant “within
-
discipline”
research. Reports for the SLCs highlight breadth enhanced by crossing
boundaries at various levels, but also depth associated with the pursuit of

16

questions within speci
fic fields. At its best, the SLC network operated on the
principle that highly specific, smaller
-
grained research both informs and is
informed by boundary
-
crossing research. Intellectual management of the SLCs
entailed finding (and then explaining to NSF

as a cooperative agreement
partner) an authentically balanced and adjustable mix of research teams and
studies. This management task in itself was seen by NSF as advancing how the
different fields in the science(s) of learning situate themselves in a cha
nging
landscape. The SLC researchers affirmed the intrinsic value to the (benignly)
coerced “between and within” integrative forces that characterized their
performance agreements with NSF.


The SLCs have also promoted a new generation of students aimed at
interdisciplinary work, who are gaining traction and success in their research
and publications. The publishing success of these early career researchers
suggests an important capacity to f
ormulate and situate strategically valuable
questions and to size up their within
-
boundary and across
-
boundary
dimensions. The (former) students present at the workshop spoke
enthusiastically about the center mode and believe that they have been well
equ
ipped to pursue interdisciplinary studies for the rest of their careers.


The centers are large entities, each incorporating many researchers from many
institutions and that scale brings with it a certain unwieldiness. They also
represent large investme
nts by NSF and that requires monitoring and
management that can be time
-
consuming, cumbersome, and expensive. Center
representatives expressed frustration at the time and money required for the
annual site visits and other reporting requirements, which we
re not always very
productive. Much of this is inevitable, given the general requirements for NSF
centers and cooperative agreements.


While the centers report a large number of publications and peer
-
reviewed
research conference presentations through SL
C support, it also needs to be said
that the they have not always been as visible in some of the most prominent
meetings and journals in several very successful areas of the science of learning,
such as language acquisition and visual development. Further
more, some of
their valuable achievements deal not with learning itself and but with the
properties of cognitive or neural functions regardless of strict matters of
development and learning. Again that reflects the large scale of the centers and
how they
are assembled in the course of submitting a proposal to NSF, drawing
together colleagues with a broad range and then struggling to generate coherent
collaboration on matters of learning. Some of the most important work in the
field of learning has not bee
n included within the centers, due in part to these
issues.


In addition, for all the emphasis on integrative approaches, it is not clear that
much has been achieved in defining a general science of learning with its own
general principles that cross the d
omains and different levels of analysis in STEM

17

learning, language acquisition, development of spatial cognition, etc, each of
which has its own distinct principles. This makes the science of learning more
like the umbrella of computer science than the ap
parently more cohesive physics
or chemistry. It has often been suggested that referring to the “sciences of
learning” rather that the singular science of learning might give a more accurate
feel for the field.


A 2009 COV found that the SLC program throug
h its six centers had succeeded in
fostering a scientific community in the science of learning, construed in this
broad fashion. This happened partly through the scale of the centers and
through creating a learning network, going beyond the individual cen
ters. The
community has been internationalized; the OECD held a workshop in January
2012 on the science of learning, where the NSF initiative played a central role,
and now the centers model is being followed and funded in other countries.
However, works
hop participants argued that
,

while the centers have clearly
been productive and

have succeeded in generating new research
, future work in
the science of learning will benefit from a more diverse set of funding
mechanisms, rather than simply funding a set
of new centers. It will be
important to continue to foster the beneficial aspects of the productive
interdisciplinarity that has been cultivated through the centers but to do it in a
way that diversifies the avenues for advancing the field and that reduce
s the
unwieldy research management problems that often characterize large center
funding.


That will be the subject

matter of the second workshop.


9
.
Challenge
s

One challenge for the future will be to remedy the absence of critical element
s of
learning fr
om the current Science of Learning C
enters program. Four

areas
where there has been good work that
has not been a focus for the centers and
will benefit from future nourishment are the following.





S
ocial/emotional/attentional factors and individual
differences
:

Current models
of learning
tend to consider only the informational
cont
ent of the input

to a generic learner.


A challenge for the future is to
understand how to model additional psychological factors that are
known to affect learning, su
ch as

social situation, emotion

and attention,
as well as individual differences in learner performance.


One difficulty
is in determining whether these factors are simply regul
ating
information uptake (e.g.
an inattentive learner just “misses” some of the
data
) or whether they qualitatively change the computations
performed.


New findings will also have implications for machine
learning: can social/emotional factors be effectively replaced by more
or different kinds of data that are easily available to machines
, or will
we need to simulate emotions and social interaction for effective
machine learning to occur?


18



Childhood acquisition of parametric variation in language capacities:

There has been a great deal of productive work on how children acquire
systematic
properties of their native language that differ from what is
found in other languages,
which differ around the world,
both in syntax
and phonology.

We are now well
positioned to focus more
intensely
on
the acquisition and “learnability” of

the parametric
variation

under

normal childhood conditions
.





How groups and societies learn:

sometime
s

societies from university
departments to nations undergo structural shifts in attitudes and
political perspectives. A striking example came around 1990, when
many countries shifted from authoritarian and totalitarian regimes to
systems where individual c
itizens had more power. This is a kind of
learning and much insightful work has been done, while mysteries
remain.



Cultural influences on learning
:

Although learning, by its vary nature, is
a culturally shaped process, as many learning scientists note, ‘
the
learning sciences have not yet adequately addressed the ways that
culture is integral to learning’ (Nasir, Rosebery, Warren & Lee 2005).
Thus, the influence and impact of culture and its variations will be as
important to identify as the impact and inf
luence of other
environmental features and variations.


10. Second workshop

The second workshop will take place in February
-
March of 2013 and will
consider
, in light of the history described in this report,

the prospects for future
work in the science of learning and how it can be best organized in terms of
funding possibilities. Speakers have been invited to address prospects for work
on memory, genetics, brain plasticity, language and cognitive developmen
t; one
or two other areas will be added. In addition, speakers will be invited to discuss
the strengths and weaknesses of various funding mechanisms, including a
national synthesis center and possible support from foundations

and other
agencies
, given the

goal of continuing to foster interdisciplinary approaches to
learning and young researchers at the graduate, postdoc and early career stages.


Submitted by the Steering Committee:

David W. Lightfoot, PI, Georgetown U

Ralph Etienne
-
Cummings, Johns Hopkins
U

Morton Gernsbacher, U Wisconsin

Eric Hamilton, Pepperdine
U

Barbar
a

Landau, Johns Hopkins U

Eliss
a Newport, Georgetown U

David Poeppel, New York U





19

References

Ananthanarayanan, R., S.K. Esser, H.D. Simon & D.S. Modha 2009
.

The cat is out
of the bag
:
Cortical simulations with 10
9

neurons, 10
13

synapses.
Proceedings of
the Conference

on High Performance Computing Networking, Storage and
Analysis
, p.63. ACM.


Carr, C. &
M.
Konishi

1988. Axonal delay lines for time measurement in the owl’s
brainstem.
PNA
S

85
:

8311
-
8315.


Chomsky
,

N. 1959.

Review of B. F. Skinner
Verbal Behavior
.
Language

35.1: 26
-
58.


Chomsky
,

N. 1965.
Aspects of the theory of syntax
. MIT Press.


Dehaene, S. 1997.


The number sense
. Oxford UP.


Dikker, S., H. Rabagliati, T.A. Farmer

& L. Pylkk
änen 2010.

Early occipital
sensitivity to syntactic category is based on form typicality.
Psychological
Science

21: 293
-
321.


Ellenbogen, J.M., J.D. Payne & R. Stickgold 2006.

The role of sleep in dec
larative
memory consolidation: P
assive, pe
rmissive, active or none?
Current Opinion in
Neurobiology

16
:

1
-
7.


Engle, R.
W.,
S.W. Tuholski, J.E. Laughlin & A.R.A.

Con
way 1999
. Working memory,
short
-
term memory, and general fluid intelligence: A latent
-
variable approach.

Journal of Experimental
Psychology: General

128:
309
-
331.


Feldman, D.E. 1999.

LTP and LTD induced by action
-
potential
-
epsp pairing at
vertical

inputs to layer II/III pyramids in rat somatosensory cortex.

Soc.
Neurosci. Abst.

25
:223.


Gallistel, C.
R. 1990.


The organization of
learning
.


MIT Press.


Gallistel, C.R., A.L. Brown, S. Carey, R. Gelman, & F.C. Keil 1991. Lessons from
animal learning for the study of cognitive development. In S. Carey & R. Gelman,
eds.
The epigenesis of mind: Essays on biology and cognition.

Lawrenc
e Erlbaum.


Grothe, B. 2003. New roles for synaptic inhibition in sound localization.
Nature
Reviews Neuroscience

4: 540
-
550.


Hermann, E., J. Call, M.V. Hernandez
-
Lloredia, B. Hare &

M. Tomasello 2009.
Humans have evolved specialized skills of social cognition: The cultural
intelligence hypothesis.
Science 317
: 1360
-
1366.


Hermer
, L

&
E.
Spelke

1996.

Modularity and development: the case of spatial

20

reorientation.
Cognition
61
:

195
-
232.

Hinton, G.
E.
,

S. Osindero & Y.
-
W. Teli 2006
.

A fast learning algorithm for deep
belief nets.
Neural Computation

18.7: 1527
-
1154.

Hopfield, J. J. 1987. Learning algorithms and probability distributions in feed
-
forward and feed
-
back networks.
PN
AS

84.23: 8429
-
8433.

Hubel D. & T. Wiesel 1962. Receptive fields, binocular interaction and functional
architecture in the cat's visual cortex.

Journal of Physiology

160: 106
-
154.


Indiveri, G., E. Chicca & R. Douglas 2006
.


VLSI array of low
-
power
spiking
neurons and bistable synapses with spike
-
timing dependent plasticity.


IEEE
Transactions o
n Neural Networks
17.1: 211
-
221.


Knudsen, E. I. 1999. Mechanisms of experience
-
dependent plasticity in the
auditory localization pathway of the barn owl.
Jou
rnal of Comparative Physiology

185
:

305

321.

Knudsen, E. 2004. Sensitive periods in the development of brain and behavior.
J.

Cognitive Neuroscience

16
:

1412
-
1425.


Lee, T.S.

& D. Mumford 2003
.

Hierarchical Bayesian
inference in the visual
cortex.

J.
Opt. Soc. Am. A
. 20.7, July 20.


Markram H., J. Lubke, M. Frotscher & B. Sakmann 1997. Regulation of synaptic

efficacy by coincidence of postsynaptic APs and EPSPs.
Science
275
: 213
-
215.


Mead C. & M.A. Mahowald 1988
.

A silicon model of early visual
processing
.
Pergamon.


Merzenich M
.
M
.
,
J.H.
Kaa
s
,
J.
Wall,
R.J.
Nelson,
M.
Su
r &

D.
Felleman 1983
.

Topographic reorganization of somatosensory cortical areas 3b and 1 in adult
monkeys following restricted deafferentation.


Neuroscience

8
.
1:

33
-
55.


Minsky,

M. & S. Papert 1969
.


Perceptrons
.

MIT Press.


Nasir, N.S., A.S. Rosebery, B. Warren & C.D. Lee 2005. Learning as a cultural
process: Achieving equity through diversity. In R.K. Sawyer, ed.
The Cambridge
Handbook of the Learning Sciences
. Cambridge UP.


Newcombe, N.S. &
K.R. Ratliff 2007.

Explaining the development of spatial
reorientation: Modularity
-
plus
-
language versus the emergence of adaptive
combination.

In

J. Plumert & J. Spencer, eds.

The emerging spatial mind
. Oxford
UP. Pp53
-
76.



21

O'Keefe, J. &

L. Nadel 1978.
The hippocampus as a cognitive map
.

Oxford UP.


Saffran, J.R., R.N. Aslin & E.L. Newport 1996.


Statistical learning by 8
-
month
-
old
infants.


Science

274: 1926

1928.


Sawyer, R.K. 2008. Optimi
zing learning: Implications of learning s
ciences
research.

From
Innovating to learn, learning to innovate
(pp45
-
61)
.


Centre for
Educational Research and Innovation.



Sejnowski, T. & C. Rosenberg 1987 Parallel networks that learn to pronounce
English text.
Complex Systems

1: 145
-
168.


Tenenbau
m, J.B., C. Kemp, T.L. Griffiths & N.D. Goodman 2011.
How to
g
row a
m
ind: Statistics,
s
tructure, and
a
bstraction.
Science
331.6022: 1279
-
1285.


Tinbergen, N. 1951.
The study of instinct
. Oxford UP, Clarendon.


Vapnik, V., S.E. Golowich & A. Smola 1997.
Support v
ect
or method for function
approximation, regression estimation, and signal processing.

Adv. Neural
Information Processing Sys
tems

9:281
-
287
.




Wang B, Y. Fan, M. Lu, S. Li, Z. Song, X. Peng, R. Zhang, Q. Lin, Y. He, J. Wang & R.
Huang 2012. Brai
n anatomical networks in world class gymnasts: A DTI
tractography study.
Neuroimage. doi: 10.1016/j.neuroimage.2012.10.007. [Epub
ahead of print]