Frontiers in Computational Neuroscience,
Год журнала:
2019,
Номер
13
Опубликована: Апрель 12, 2019
Deep
neural
networks
(DNNs)
have
recently
been
applied
successfully
to
brain
decoding
and
image
reconstruction
from
functional
magnetic
resonance
imaging
(fMRI)
activity.
However,
direct
training
of
a
DNN
with
fMRI
data
is
often
avoided
because
the
size
available
thought
be
insufficient
for
complex
network
numerous
parameters.
Instead,
pre-trained
usually
serves
as
proxy
hierarchical
visual
representations,
are
used
decode
individual
features
stimulus
using
simple
linear
model,
which
then
passed
module.
Here,
we
directly
trained
model
corresponding
images
build
an
end-to-end
model.
We
accomplished
this
by
generative
adversarial
additional
loss
term
that
was
defined
in
high-level
feature
space
(feature
loss)
up
6,000
samples
(natural
responses).
The
above
tested
on
independent
datasets
reconstructed
pattern
input.
Reconstructions
obtained
our
proposed
method
resembled
test
stimuli
artificial
images)
accuracy
increased
function
training-data
size.
Ablation
analyses
indicated
employed
played
critical
role
achieving
accurate
reconstruction.
Our
results
show
can
learn
mapping
between
activity
perception.
IEEE Transactions on Knowledge and Data Engineering,
Год журнала:
2020,
Номер
34(8), С. 3681 - 3700
Опубликована: Сен. 23, 2020
With
the
fast
development
of
various
positioning
techniques
such
as
Global
Position
System
(GPS),
mobile
devices
and
remote
sensing,
spatio-temporal
data
has
become
increasingly
available
nowadays.
Mining
valuable
knowledge
from
is
critically
important
to
many
real-world
applications
including
human
mobility
understanding,
smart
transportation,
urban
planning,
public
safety,
health
care
environmental
management.
As
number,
volume
resolution
increase
rapidly,
traditional
mining
methods,
especially
statistics-based
methods
for
dealing
with
are
becoming
overwhelmed.
Recently
deep
learning
models
recurrent
neural
network
(RNN)
convolutional
(CNN)
have
achieved
remarkable
success
in
domains
due
powerful
ability
automatic
feature
representation
learning,
also
widely
applied
(STDM)
tasks
predictive
anomaly
detection
classification.
In
this
paper,
we
provide
a
comprehensive
review
recent
progress
applying
STDM.
We
first
categorize
into
five
different
types,
then
briefly
introduce
that
used
Next,
classify
existing
literature
based
on
types
data,
tasks,
models,
followed
by
STDM
on-demand
service,
climate
&
weather
analysis,
mobility,
location-based
social
network,
crime
neuroscience.
Finally,
conclude
limitations
current
research
point
out
future
directions.
Cerebral Cortex,
Год журнала:
2017,
Номер
28(12), С. 4136 - 4160
Опубликована: Сен. 23, 2017
Convolutional
neural
network
(CNN)
driven
by
image
recognition
has
been
shown
to
be
able
explain
cortical
responses
static
pictures
at
ventral-stream
areas.
Here,
we
further
showed
that
such
CNN
could
reliably
predict
and
decode
functional
magnetic
resonance
imaging
data
from
humans
watching
natural
movies,
despite
its
lack
of
any
mechanism
account
for
temporal
dynamics
or
feedback
processing.
Using
separate
data,
encoding
decoding
models
were
developed
evaluated
describing
the
bi-directional
relationships
be-tween
brain.
Through
models,
CNN-predicted
areas
covered
not
only
ventral
stream,
but
also
dorsal
albe-it
a
lesser
degree;
single-voxel
response
was
visualized
as
specific
pixel
pattern
drove
response,
revealing
distinct
representation
individual
location;
activation
synthesized
images
with
high-throughput
map
category
representation,
con-trast,
selectivity.
fMRI
signals
directly
decoded
estimate
feature
representations
in
both
visual
semantic
spaces,
direct
reconstruction
seman-tic
categorization,
respectively.
These
results
cor-roborate,
generalize,
extend
previous
findings,
highlight
value
using
deep
learning,
an
all-in-one
model
cortex,
understand
vision.
Annual Review of Vision Science,
Год журнала:
2019,
Номер
5(1), С. 399 - 426
Опубликована: Авг. 8, 2019
Artificial
vision
has
often
been
described
as
one
of
the
key
remaining
challenges
to
be
solved
before
machines
can
act
intelligently.
Recent
developments
in
a
branch
machine
learning
known
deep
have
catalyzed
impressive
gains
vision-giving
sense
that
problem
is
getting
closer
being
solved.
The
goal
this
review
provide
comprehensive
overview
recent
and
critically
assess
actual
progress
toward
achieving
human-level
visual
intelligence.
I
discuss
implications
successes
limitations
modern
algorithms
for
biological
prospect
neuroscience
inform
design
future
artificial
systems.
PLoS Computational Biology,
Год журнала:
2019,
Номер
15(1), С. e1006633 - e1006633
Опубликована: Янв. 14, 2019
The
mental
contents
of
perception
and
imagery
are
thought
to
be
encoded
in
hierarchical
representations
the
brain,
but
previous
attempts
visualize
perceptual
have
failed
capitalize
on
multiple
levels
hierarchy,
leaving
it
challenging
reconstruct
internal
imagery.
Recent
work
showed
that
visual
cortical
activity
measured
by
functional
magnetic
resonance
imaging
(fMRI)
can
decoded
(translated)
into
features
a
pre-trained
deep
neural
network
(DNN)
for
same
input
image,
providing
way
make
use
information
from
features.
Here,
we
present
novel
image
reconstruction
method,
which
pixel
values
an
optimized
its
DNN
similar
those
human
brain
at
layers.
We
found
our
method
was
able
reliably
produce
reconstructions
resembled
viewed
natural
images.
A
prior
introduced
generator
effectively
rendered
semantically
meaningful
details
reconstructions.
Human
judgment
supported
effectiveness
combining
layers
enhance
quality
generated
While
model
solely
trained
with
images,
successfully
generalized
artificial
shapes,
indicating
not
simply
matching
exemplars.
analysis
applied
demonstrated
rudimentary
subjective
content.
Our
results
suggest
combine
new
window
brain.
