bioRxiv (Cold Spring Harbor Laboratory),
Journal Year:
2023,
Volume and Issue:
unknown
Published: Aug. 6, 2023
The
activity
of
single
neurons
encodes
behavioral
variables,
such
as
sensory
stimuli
(Hubel
&
Wiesel
1959)
and
choice
(Britten
et
al.
1992;
Guo
2014),
but
their
influence
on
behavior
is
often
mysterious.
We
estimated
the
a
unit
neural
from
recordings
in
anterior
lateral
motor
cortex
(ALM)
mice
performing
memory-guided
movement
task
(H.
K.
Inagaki
2018).
Choice
selectivity
grew
it
flowed
through
sequence
directions
space.
Early
carried
little
were
predicted
to
have
large
influence,
while
late
influence.
Consequently,
was
only
weakly
correlated
with
selectivity;
proportion
selective
for
one
opposite
direction.
These
results
consistent
models
which
recurrent
circuits
produce
feedforward
amplification
(Goldman
2009;
Ganguli
2008;
Murphy
Miller
2009)
so
that
small
amplitude
signals
along
early
are
amplified
low-dimensional
directions,
behavior.
Targeted
photostimulation
experiments
(Daie
2021b)
revealed
triggered
sequential
later
caused
predictable
biases.
demonstrate
existence
an
amplifying
dynamical
motif
cortex,
explain
paradoxical
responses
perturbation
(Chettih
Harvey
2019;
Daie
2021b;
Russell
2019),
reveal
relevance
dynamics.
Current Opinion in Neurobiology,
Journal Year:
2021,
Volume and Issue:
70, P. 137 - 144
Published: Oct. 1, 2021
Advances
in
experimental
neuroscience
have
transformed
our
ability
to
explore
the
structure
and
function
of
neural
circuits.
At
same
time,
advances
machine
learning
unleashed
remarkable
computational
power
artificial
networks
(ANNs).
While
these
two
fields
different
tools
applications,
they
present
a
similar
challenge:
namely,
understanding
how
information
is
embedded
processed
through
high-dimensional
representations
solve
complex
tasks.
One
approach
addressing
this
challenge
utilize
mathematical
analyze
geometry
representations,
i.e.,
population
geometry.
We
review
examples
geometrical
approaches
providing
insight
into
biological
networks:
representation
untangling
perception,
geometric
theory
classification
capacity,
disentanglement
abstraction
cognitive
systems,
topological
underlying
maps,
dynamic
motor
dynamical
cognition.
Together,
findings
illustrate
an
exciting
trend
at
intersection
learning,
neuroscience,
geometry,
which
provides
useful
population-level
mechanistic
descriptor
task
implementation.
Importantly,
descriptions
are
applicable
across
sensory
modalities,
brain
regions,
network
architectures
timescales.
Thus,
has
potential
unify
networks,
bridging
gap
between
single
neurons,
populations
behavior.
Nature Communications,
Journal Year:
2021,
Volume and Issue:
12(1)
Published: March 3, 2021
Abstract
Artificial
neural
networks
have
recently
achieved
many
successes
in
solving
sequential
processing
and
planning
tasks.
Their
success
is
often
ascribed
to
the
emergence
of
task’s
low-dimensional
latent
structure
network
activity
–
i.e.,
learned
representations.
Here,
we
investigate
hypothesis
that
a
means
for
generating
representations
with
easily
accessed
structure,
possibly
reflecting
an
underlying
semantic
organization,
through
learning
predict
observations
about
world.
Specifically,
ask
whether
when
mechanisms
sensory
prediction
coincide
those
extracting
variables.
Using
recurrent
model
trained
sequence
show
dynamics
exhibit
but
nonlinearly
transformed
inputs
map
environment.
We
quantify
these
results
using
nonlinear
measures
intrinsic
dimensionality
linear
decodability
variables,
provide
mathematical
arguments
why
such
useful
predictive
emerge.
focus
throughout
on
how
our
can
aid
analysis
interpretation
experimental
data.
