Nature Communications,
Год журнала:
2025,
Номер
16(1)
Опубликована: Янв. 22, 2025
Artificial
neural
networks
(ANNs)
are
at
the
core
of
most
Deep
Learning
(DL)
algorithms
that
successfully
tackle
complex
problems
like
image
recognition,
autonomous
driving,
and
natural
language
processing.
However,
unlike
biological
brains
who
similar
in
a
very
efficient
manner,
DL
require
large
number
trainable
parameters,
making
them
energy-intensive
prone
to
overfitting.
Here,
we
show
new
ANN
architecture
incorporates
structured
connectivity
restricted
sampling
properties
dendrites
counteracts
these
limitations.
We
find
dendritic
ANNs
more
robust
overfitting
match
or
outperform
traditional
on
several
classification
tasks
while
using
significantly
fewer
parameters.
These
advantages
likely
result
different
learning
strategy,
whereby
nodes
respond
multiple
classes,
classical
strive
for
class-specificity.
Our
findings
suggest
incorporation
can
make
precise,
resilient,
parameter-efficient
shed
light
how
features
impact
strategies
ANNs.
IEEE Signal Processing Magazine,
Год журнала:
2019,
Номер
36(6), С. 51 - 63
Опубликована: Ноя. 1, 2019
Spiking
neural
networks
(SNNs)
are
nature's
versatile
solution
to
fault-tolerant,
energy-efficient
signal
processing.
To
translate
these
benefits
into
hardware,
a
growing
number
of
neuromorphic
spiking
NN
processors
have
attempted
emulate
biological
NNs.
These
developments
created
an
imminent
need
for
methods
and
tools
that
enable
such
systems
solve
real-world
processing
problems.
Like
conventional
NNs,
SNNs
can
be
trained
on
real,
domain-specific
data;
however,
their
training
requires
the
overcoming
challenges
linked
binary
dynamical
nature.
This
article
elucidates
step-by-step
problems
typically
encountered
when
guides
reader
through
key
concepts
synaptic
plasticity
data-driven
learning
in
setting.
Accordingly,
it
gives
overview
existing
approaches
provides
introduction
surrogate
gradient
(SG)
methods,
specifically,
as
particularly
flexible
efficient
method
overcome
aforementioned
challenges.
Advanced Materials,
Год журнала:
2019,
Номер
31(49)
Опубликована: Сен. 24, 2019
As
the
research
on
artificial
intelligence
booms,
there
is
broad
interest
in
brain-inspired
computing
using
novel
neuromorphic
devices.
The
potential
of
various
emerging
materials
and
devices
for
has
attracted
extensive
efforts,
leading
to
a
large
number
publications.
Going
forward,
order
better
emulate
brain's
functions,
its
relevant
fundamentals,
working
mechanisms,
resultant
behaviors
need
be
re-visited,
understood,
connected
electronics.
A
systematic
overview
biological
neural
systems
given,
along
with
their
related
critical
mechanisms.
Recent
progress
reviewed
and,
more
importantly,
existing
challenges
are
highlighted
hopefully
shed
light
future
directions.
Annual Review of Neuroscience,
Год журнала:
2020,
Номер
43(1), С. 95 - 117
Опубликована: Фев. 20, 2020
Synaptic
plasticity,
the
activity-dependent
change
in
neuronal
connection
strength,
has
long
been
considered
an
important
component
of
learning
and
memory.
Computational
engineering
work
corroborate
power
through
directed
adjustment
weights.
Here
we
review
fundamental
elements
four
broadly
categorized
forms
synaptic
plasticity
discuss
their
functional
capabilities
limitations.
Although
standard,
correlation-based,
Hebbian
primary
focus
neuroscientists
for
decades,
it
is
inherently
limited.
Three-factor
rules
supplement
with
neuromodulation
eligibility
traces,
while
true
supervised
types
go
even
further
by
adding
objectives
instructive
signals.
Finally,
a
recently
discovered
hippocampal
form
combines
above
elements,
leaving
behind
requirement.
We
suggest
that
effort
to
determine
neural
basis
adaptive
behavior
could
benefit
from
renewed
experimental
theoretical
investigation
more
powerful
plasticity.
Frontiers in Neuroscience,
Год журнала:
2018,
Номер
12
Опубликована: Окт. 25, 2018
Spiking
neural
networks
(SNNs)
are
inspired
by
information
processing
in
biology,
where
sparse
and
asynchronous
binary
signals
communicated
processed
a
massively
parallel
fashion.
SNNs
on
neuromorphic
hardware
exhibit
favorable
properties
such
as
low
power
consumption,
fast
inference,
event-driven
processing.
This
makes
them
interesting
candidates
for
the
efficient
implementation
of
deep
networks,
method
choice
many
machine
learning
tasks.
In
this
review,
we
address
opportunities
that
spiking
offer
investigate
detail
challenges
associated
with
training
way
competitive
conventional
learning,
but
simultaneously
allows
mapping
to
hardware.
A
wide
range
methods
is
presented,
ranging
from
conversion
into
SNNs,
constrained
before
conversion,
variants
backpropagation,
biologically
motivated
STDP.
The
goal
our
review
define
categorization
SNN
methods,
summarize
their
advantages
drawbacks.
We
further
discuss
relationships
between
which
becoming
popular
digital
implementation.
Neuromorphic
platforms
have
great
potential
enable
real-world
applications.
compare
suitability
various
systems
been
developed
over
past
years,
use
cases.
approaches
should
not
be
considered
simply
two
solutions
same
classes
problems,
instead
it
possible
identify
exploit
task-specific
advantages.
Deep
work
new
types
event-based
sensors,
temporal
codes
local
on-chip
so
far
just
scratched
surface
realizing
these
practical
Neural Computation,
Год журнала:
2018,
Номер
30(6), С. 1514 - 1541
Опубликована: Апрель 13, 2018
A
vast
majority
of
computation
in
the
brain
is
performed
by
spiking
neural
networks.
Despite
ubiquity
such
spiking,
we
currently
lack
an
understanding
how
biological
circuits
learn
and
compute
vivo,
as
well
can
instantiate
capabilities
artificial
silico.
Here
revisit
problem
supervised
learning
temporally
coding
multilayer
First,
using
a
surrogate
gradient
approach,
derive
SuperSpike,
nonlinear
voltage-based
three-factor
rule
capable
training
networks
deterministic
integrate-and-fire
neurons
to
perform
computations
on
spatiotemporal
spike
patterns.
Second,
inspired
recent
results
feedback
alignment,
compare
performance
our
under
different
credit
assignment
strategies
for
propagating
output
errors
hidden
units.
Specifically,
test
uniform,
symmetric,
random
feedback,
finding
that
simpler
tasks
be
solved
with
any
type
while
more
complex
require
symmetric
feedback.
In
summary,
open
door
obtaining
better
scientific
advancing
ability
train
them
solve
problems
involving
transformations
between
time
Trends in Cognitive Sciences,
Год журнала:
2019,
Номер
23(3), С. 235 - 250
Опубликована: Янв. 30, 2019
The
error
back-propagation
algorithm
can
be
approximated
in
networks
of
neurons,
which
plasticity
only
depends
on
the
activity
presynaptic
and
postsynaptic
neurons.
These
biologically
plausible
deep
learning
models
include
both
feedforward
feedback
connections,
allowing
errors
made
by
network
to
propagate
through
layers.
rules
different
implemented
with
types
spike-time-dependent
plasticity.
dynamics
described
within
a
common
framework
energy
minimisation.
This
review
article
summarises
recently
proposed
theories
how
neural
circuits
brain
could
approximate
used
artificial
networks.
Computational
implementing
these
achieve
as
efficient
networks,
but
they
use
simple
synaptic
based
have
similarities,
such
including
information
about
throughout
network.
Furthermore,
incorporate
experimental
evidence
connectivity,
responses,
provide
insights
might
organised
that
modification
weights
multiple
levels
cortical
hierarchy
leads
improved
performance
tasks.
In
past
few
years,
computer
programs
using
(see
Glossary)
achieved
impressive
results
complex
cognitive
tasks
were
previously
reach
humans.
processing
natural
images
language
[1LeCun
Y.
et
al.Deep
learning.Nature.
2015;
521:
436-444Crossref
PubMed
Scopus
(42113)
Google
Scholar],
or
playing
arcade
board
games
[2Mnih
V.
al.Human-level
control
reinforcement
518:
529-533Crossref
(13741)
Scholar,
3Silver
D.
al.Mastering
game
Go
tree
search.Nature.
2016;
529:
484-489Crossref
(8554)
Scholar].
Since
recent
applications
extended
versions
classic
[4Rumelhart
D.E.
al.Learning
representations
back-propagating
errors.Nature.
1986;
323:
533-536Crossref
(15380)
their
success
has
inspired
studies
comparing
brain.
It
been
demonstrated
when
learn
perform
image
classification
navigation,
neurons
layers
develop
similar
those
seen
areas
involved
tasks,
receptive
fields
across
visual
grid
cells
entorhinal
cortex
[5Banino
A.
al.Vector-based
navigation
grid-like
agents.Nature.
2018;
557:
429-433Crossref
(289)
6Whittington,
J.C.R.
al.
(2018)
Generalisation
structural
knowledge
hippocampal-entorhinal
system.
31st
Conference
Neural
Information
Processing
Systems
(NIPS
2018),
MontrealGoogle
7Yamins
D.L.
DiCarlo
J.J.
Using
goal-driven
understand
sensory
cortex.Nat.
Neurosci.
19:
356-365Crossref
(650)
suggests
may
analogous
algorithms.
thanks
current
computational
advances,
now
useful
functions
are
[8Bowers
J.S.
Parallel
distributed
theory
age
networks.Trends
Cogn.
Sci.
2017;
21:
950-961Abstract
Full
Text
PDF
(22)
A
key
question
remains
open
is
implement
describes
connections
should
modified
during
learning,
its
attractiveness,
part,
comes
from
prescribing
weight
changes
reduce
network,
according
theoretical
analysis.
Although
originally
brain,
weights,
appears
unrealistic
[9Crick
F.
excitement
networks.Nature.
1989;
337:
129-132Crossref
(353)
10Grossberg
S.
Competitive
learning:
interactive
activation
adaptive
resonance.Cogn.
1987;
11:
23-63Crossref
Nevertheless,
[11Bengio
al.STDP-Compatible
approximation
backpropagation
an
energy-based
model.Neural
Comput.
29:
555-577Crossref
(47)
12Guerguiev
J.
al.Towards
segregated
dendrites.eLife.
6e22901Crossref
(173)
13Sacramento,
Dendritic
microcircuits
algorithm.
14Whittington
Bogacz
R.
An
predictive
coding
local
Hebbian
plasticity.Neural
1229-1262Crossref
(91)
theoretic
important
because
overrule
dogma,
generally
accepted
for
30
too
complicated
Before
discussing
this
new
generation
detail,
we
first
brief
overview
train
discuss
why
it
was
considered
implausible.
To
effectively
feedback,
often
need
appropriately
adjusted
hierarchical
simultaneously.
For
example,
child
learns
name
letters,
incorrect
pronunciation
combined
result
speech,
associative,
areas.
When
multi-layer
makes
error,
assigns
credit
individual
synapses
all
prescribes
much.
How
networks?
trained
set
examples,
each
consisting
input
pattern
target
pattern.
pair,
generates
prediction
then
minimise
difference
between
predicted
determine
appropriate
modification,
term
computed
neuron
change
discrepancy
(Box
1).
Each
amount
determined
product
projects
to.Box
1Artificial
NetworksA
conventional
consists
layer
receiving
weighted
previous
(Figure
IA).
propagating
layers,
Equation
1.1,
where
xl
vector
denoting
l
Wl−1
matrix
−
1
l.
function
f
applied
allow
nonlinear
computations.During
cost
quantifying
patterns
(typically
defined
1.2).
particular,
direction
steepest
decrease
(or
gradient)
ID).
Such
1.3,
δl+1
terms
associated
xl+1.
last
L
1.4
t
activity.
Thus,
output
positive
if
higher
than
earlier
1.5
sum
above
strengths
(and
further
scaled
derivative
function;
·
denotes
element-wise
multiplication).
hidden
unit
sends
excitatory
projections
units
high
terms,
so
increasing
would
output.
Once
computed,
changed
1.3
proportion
neuron.
computations.
During
procedure
steps
take
place
case
naming
letters
mentioned
above,
corresponds
letter.
After
seeing
image,
guess
at
(predicted
pattern)
via
speech
On
supervision
his
her
parent
correct
(target
pattern),
along
stream
more
likely
sound
will
produced
again.
algorithmic
process
enough,
there
problems
biology.
Below,
briefly
three
issues.
Conventional
compute
forward
direction,
separately
external
Without
representation,
update
computations
downstream
biological
connection
strength
solely
signals
(e.g.,
connect),
unclear
afforded
Historically,
major
criticism;
thus
main
focus
our
article.
back-propagated
same
prediction.
symmetry
identical
exist
directions
connected
bidirectional
significantly
expected
chance,
not
always
present
[15Song
al.Highly
nonrandom
features
connectivity
circuits.PLoS
Biol.
2005;
3:
507-519Google
even
present,
backwards
forwards
still
correctly
align
themselves.
Artificial
send
continuous
(corresponding
firing
rate
neurons),
whereas
real
spikes.
Generalising
discrete
spikes
trivial,
derivate
found
Away
algorithm,
description
inside
also
simplified
linear
summation
inputs.
above-mentioned
issues
investigated
studies.
lack
representation
addressed
early
proposing
instead
driven
global
signal
carried
neuromodulators
[16Mazzoni
P.
al.A
rule
networks.Proc.
Natl.
Acad.
U.
1991;
88:
4433-4437Crossref
(138)
17Williams
R.J.
Simple
statistical
gradient-following
algorithms
connectionist
learning.Mach.
Learn.
1992;
8:
229-256Crossref
18Unnikrishnan
K.P.
Venugopal
Alopex:
correlation-based
recurrent
networks.Neural
1994;
6:
469-490Crossref
19Seung
H.S.
Learning
spiking
stochastic
transmission.Neuron.
2003;
40:
1063-1073Abstract
(238)
However,
slow
does
scale
size
[20Werfel
curves
gradient
descent
17:
2699-2718Crossref
More
promisingly,
several
do
represent
locally
closely
similarly
standard
benchmark
handwritten
digit
classification)
[12Guerguiev
21Lillicrap
T.P.
al.Random
support
learning.Nat.
Commun.
713276Crossref
(336)
22Scellier
B.
Bengio
Equilibrium
propagation:
bridging
gap
backpropagation.Front.
24Crossref
(146)
summarise
them
detail
following
sections.
criticism
demonstrating
random
good
[21Lillicrap
23Zenke
Ganguli
SuperSpike:
supervised
multilayer
30:
1514-1541Crossref
(209)
24Mostafa,
H.
(2017)
Deep
errors.
arXiv
preprint
arXiv:1711.06756Google
25Scellier,
Generalization
equilibrium
propagation
field
dynamics.
1808.04873Google
26Liao,
Q.
(2016)
backpropagation?
AAAI
Intelligence,
pp.
1837–1844,
AAAIGoogle
27Baldi
Sadowski
channel,
optimality
backpropagation.Neural
Netw.
83:
51-74Crossref
(39)
being
said,
some
concern
regarding
issue
[28Bartunov,
Assessing
scalability
biologically-motivated
architectures.
With
regard
realism
shown
generalised
producing
[29Sporea
I.
Grüning
Supervised
2013;
25:
473-509Crossref
(97)
Scholar]
calculating
derivatives
overcome
[23Zenke
realistic
considered,
themselves
small
dendritic
structures
[30Schiess
M.
al.Somato-dendritic
error-backpropagation
active
dendrites.PLoS
12e1004638Crossref
(43)
There
diversity
ideas
[31Balduzzi,
(2015)
Kickback
cuts
backprop's
red-tape:
assignment
485–491,
32Krotov,
Hopfield,
Unsupervised
competing
units.
arXiv:1806.10181Google
33Kuśmierz
Ł.
factors:
modulating
errors.Curr.
Opin.
Neurobiol.
46:
170-177Crossref
(52)
34Marblestone
A.H.
al.Toward
integration
neuroscience.Front.
10:
94Crossref
(316)
35Bengio,
(2014)
auto-encoders
propagation.
arXiv:1407.7906Google
36Lee,
D.-H.
Difference
Joint
European
Machine
Knowledge
Discovery
Databases,
498–515,
SpringerGoogle
Scholar];
however,
principles
behind
related
37O'Reilly
R.C.
Biologically
error-driven
differences:
generalized
recirculation
algorithm.Neural
1996;
895-938Crossref
(211)
substantial
data
while
paralleling
operate
minimal
control,
modifications
depend
biology,
spike
time-dependent
plasticity,
properties
pyramidal
microcircuits.
We
emphasise
rely
fundamentally
principles.
thereby
without
requiring
program
dynamics,
well
divide
reviewed
two
classes
differing
represented,
class
model
encodes
differences
time.
contrastive
[37O'Reilly
relies
observation
proportional
(difference
decomposed
into
separate
updates:
one
other
provided
[38Ackley
D.H.
Boltzmann
machines.Cogn.
1985;
9:
147-169Crossref
2).
twice:
anti-Hebbian
once
converges
(after
propagated
connections)
role
'unlearn'
existing
association
prediction,
second
target.Box
2Temporal-Error
ModelsTemporal-error
describe
nodes
given
node
summed
inputs
adjacent
decay
level
IB).
As
recurrent,
no
longer
possible
write
equation
describing
(such
1.1
Box
1);
instead,
differential
2.1
[72Pineda
F.J.
networks.Phys.
Rev.
Lett.
59:
2229-2232Crossref
(594)
x˙l
over
time
(all
equations
figure
ignore
nonlinearities
brevity).In
model,
occurring
times.
easiest
consider
connecting
modified.
Substituting
see
2.2
required
terms.
O'Reilly
presence
backward
propagates
sequence
approximates
version
Scholar].In
gradually
(x3|¬t)
towards
values
(t),
sample
Figure
ID.
temporal
(x˙3)
(t
−x3|¬t),
is,
(defined
1.4).
Hence,
simply
equal
(Equation
2.3).
Temporal-error
brevity).
o