ACM Computing Surveys,
Journal Year:
2022,
Volume and Issue:
55(12), P. 1 - 49
Published: Nov. 17, 2022
Neuromorphic
Computing,
a
concept
pioneered
in
the
late
1980s,
is
receiving
lot
of
attention
lately
due
to
its
promise
reducing
computational
energy,
latency,
as
well
learning
complexity
artificial
neural
networks.
Taking
inspiration
from
neuroscience,
this
interdisciplinary
field
performs
multi-stack
optimization
across
devices,
circuits,
and
algorithms
by
providing
an
end-to-end
approach
achieving
brain-like
efficiency
machine
intelligence.
On
one
side,
neuromorphic
computing
introduces
new
algorithmic
paradigm,
known
Spiking
Neural
Networks
(SNNs),
which
significant
shift
standard
deep
transmits
information
spikes
(“1”
or
“0”)
rather
than
analog
values.
This
has
opened
up
novel
research
directions
formulate
methods
represent
data
spike-trains,
develop
neuron
models
that
can
process
over
time,
design
for
event-driven
dynamical
systems,
engineer
network
architectures
amenable
sparse,
asynchronous,
achieve
lower
power
consumption.
other
parallel
thrust
focuses
on
development
efficient
platforms
algorithms.
Standard
accelerators
are
workloads
not
particularly
suitable
handle
processing
multiple
timesteps
efficiently.
To
effect,
researchers
have
designed
hardware
rely
sparse
computations
matrix
operations.
While
most
large-scale
systems
been
explored
based
CMOS
technology,
recently,
Non-Volatile
Memory
(NVM)
technologies
show
toward
implementing
bio-mimetic
functionalities
single
devices.
In
article,
we
outline
several
strides
spiking
networks
(SNNs)
taken
recent
past,
present
our
outlook
challenges
needs
overcome
make
bio-plausibility
route
successful
one.
Frontiers in Neuroscience,
Journal Year:
2018,
Volume and Issue:
12
Published: Oct. 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
Frontiers in Neuroscience,
Journal Year:
2020,
Volume and Issue:
14
Published: Feb. 28, 2020
Spiking
Neural
Networks
(SNNs)
have
recently
emerged
as
a
prominent
neural
computing
paradigm.
However,
the
typical
shallow
SNN
architectures
limited
capacity
for
expressing
complex
representations,
while
training
deep
SNNs
using
input
spikes
has
not
been
successful
so
far.
Diverse
methods
proposed
to
get
around
this
issue
such
converting
off-line
trained
Artificial
(ANNs)
SNNs.
ANN-SNN
conversion
scheme
fails
capture
temporal
dynamics
of
spiking
system.
On
other
hand,
it
is
still
difficult
problem
directly
train
spike
events
due
discontinuous,
non-differentiable
nature
generation
function.
To
overcome
problem,
we
propose
an
approximate
derivative
method
that
accounts
leaky
behavior
LIF
neurons.
This
enables
convolutional
(with
events)
spike-based
backpropagation.
Our
experiments
show
effectiveness
learning
strategy
on
networks
(VGG
and
Residual
architectures)
by
achieving
best
classification
accuracies
in
MNIST,
SVHN
CIFAR-10
datasets
compared
with
learning.
Moreover,
analyze
sparse
event-based
computations
demonstrate
efficacy
inference
operation
domain.
2021 IEEE/CVF International Conference on Computer Vision (ICCV),
Journal Year:
2021,
Volume and Issue:
unknown, P. 2641 - 2651
Published: Oct. 1, 2021
Spiking
Neural
Networks
(SNNs)
have
attracted
enormous
research
interest
due
to
temporal
information
processing
capability,
low
power
consumption,
and
high
biological
plausibility.
However,
the
formulation
of
efficient
high-performance
learning
algorithms
for
SNNs
is
still
challenging.
Most
existing
methods
learn
weights
only,
require
manual
tuning
membrane-related
parameters
that
determine
dynamics
a
single
spiking
neuron.
These
are
typically
chosen
be
same
all
neurons,
which
limits
diversity
neurons
thus
expressiveness
resulting
SNNs.
In
this
paper,
we
take
inspiration
from
observation
different
across
brain
regions,
propose
training
algorithm
capable
not
only
synaptic
but
also
membrane
time
constants
We
show
incorporating
learnable
can
make
network
less
sensitive
initial
values
speed
up
learning.
addition,
reevaluate
pooling
in
find
max-pooling
will
lead
significant
loss
advantage
computation
cost
binary
compatibility.
evaluate
proposed
method
image
classification
tasks
on
both
traditional
static
MNIST,
Fashion-MNIST,
CIFAR-10
datasets,
neuromorphic
N-MNIST,
CIFAR10-DVS,
DVS128
Gesture
datasets.
The
experiment
results
outperforms
state-of-the-art
accuracy
nearly
using
fewer
time-steps.
Our
codes
available
at
https://github.com/fangwei123456/Parametric-Leaky-Integrate-and-Fire-Spiking-Neuron.
Brain Sciences,
Journal Year:
2022,
Volume and Issue:
12(7), P. 863 - 863
Published: June 30, 2022
The
past
decade
has
witnessed
the
great
success
of
deep
neural
networks
in
various
domains.
However,
are
very
resource-intensive
terms
energy
consumption,
data
requirements,
and
high
computational
costs.
With
recent
increasing
need
for
autonomy
machines
real
world,
e.g.,
self-driving
vehicles,
drones,
collaborative
robots,
exploitation
those
applications
been
actively
investigated.
In
applications,
efficiencies
especially
important
because
real-time
responses
limited
supply.
A
promising
solution
to
these
previously
infeasible
recently
given
by
biologically
plausible
spiking
networks.
Spiking
aim
bridge
gap
between
neuroscience
machine
learning,
using
realistic
models
neurons
carry
out
computation.
Due
their
functional
similarity
biological
network,
can
embrace
sparsity
found
biology
highly
compatible
with
temporal
code.
Our
contributions
this
work
are:
(i)
we
give
a
comprehensive
review
theories
neurons;
(ii)
present
existing
spike-based
neuron
models,
which
have
studied
neuroscience;
(iii)
detail
synapse
models;
(iv)
provide
artificial
networks;
(v)
detailed
guidance
on
how
train
(vi)
revise
available
frameworks
that
developed
support
implementing
(vii)
finally,
cover
network
computer
vision
robotics
paper
concludes
discussions
future
perspectives.
2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),
Journal Year:
2020,
Volume and Issue:
unknown
Published: June 1, 2020
Spiking
Neural
Networks
(SNNs)
have
recently
attracted
significant
research
interest
as
the
third
generation
of
artificial
neural
networks
that
can
enable
low-power
event-driven
data
analytics.
The
best
performing
SNNs
for
image
recognition
tasks
are
obtained
by
converting
a
trained
Analog
Network
(ANN),
consisting
Rectified
Linear
Units
(ReLU),
to
SNN
composed
integrate-and-fire
neurons
with
"proper"
firing
thresholds.
converted
typically
incur
loss
in
accuracy
compared
provided
original
ANN
and
require
sizable
number
inference
time-steps
achieve
accuracy.
We
find
performance
degradation
stems
from
using
"hard
reset"
spiking
neuron
is
driven
fixed
reset
potential
once
its
membrane
exceeds
threshold,
leading
information
during
inference.
propose
ANN-SNN
conversion
"soft
model,
referred
Residual
Membrane
Potential
(RMP)
neuron,
which
retains
"residual"
above
threshold
at
instants.
demonstrate
near
loss-less
RMP
VGG-16,
ResNet-20,
ResNet-34
on
challenging
datasets
including
CIFAR-10
(93.63%
top-1),
CIFAR-100
(70.93%
ImageNet
(73.09%
top-1
accuracy).
Our
results
also
show
RMP-SNN
surpasses
2-8
times
fewer
across
network
architectures
datasets.
Frontiers in Neuroinformatics,
Journal Year:
2018,
Volume and Issue:
12
Published: Dec. 12, 2018
The
development
of
spiking
neural
network
simulation
software
is
a
critical
component
enabling
the
modeling
systems
and
biologically
inspired
algorithms.
Existing
frameworks
support
wide
range
functionality,
abstraction
levels,
hardware
devices,
yet
are
typically
not
suitable
for
rapid
prototyping
or
application
to
problems
in
domain
machine
learning.
In
this
paper,
we
describe
new
Python
package
networks,
specifically
geared
towards
learning
reinforcement
Our
software,
called
\texttt{BindsNET},
enables
building
networks
features
user-friendly,
concise
syntax.
\texttt{BindsNET}
built
on
\texttt{PyTorch}
deep
library,
facilitating
implementation
fast
CPU
GPU
computational
platforms.
Moreover,
framework
can
be
adjusted
utilize
other
existing
computing
backends;
e.g.,
\texttt{TensorFlow}
\texttt{SpiNNaker}.
We
provide
an
interface
with
OpenAI
\texttt{gym}
allowing
training
evaluation
environments.
argue
that
facilitates
use
large-scale
show
some
simple
examples
by
using
practice.
\blfootnote{\texttt{BindsNET}
code
available
at
\texttt{https://github.com/Hananel-Hazan/bindsnet}.
To
install
version
used
\texttt{pip
bindsnet=0.2.1}.}