Elena
Voss
PhD
Applicant
·
Computational
Neuroscience
elena
.
voss
@
umich
.
edu
· (734) 555-0182
STAT EMENT
OF
PURPOSE
In
the
winter
of
2019,
I
spent
three
weeks
in
a
darkened
electrophysiology
suite
at
the
University
of
Michigan
watching
a
rhesus
macaque
named
Juno
trace
a
cursor
toward
drifting
targets
on
a
screen
.
I
was
not
there
for
the
task
itself
;
I
was
there
to
understand
how
ninety
-
six
channels
of
spiking
activity
from
a
chronically
implanted
Utah
array
could
be
translated
into
the
intention
to
reach
.
Watching
real
-
time
spike
trains
resolve
into
kinematic
trajectories
on
an
adjacent
monitor
,
I
realized
that
the
brain
is
not
an
inscrutable
black
box
but
a
dynamical
system
whose
algorithms
we
can
decode
,
model
,
and
ultimately
repair
.
That
experience
set
the
trajectory
of
my
intellectual
life
and
continues
to
animate
my
desire
to
pursue
a
PhD
in
the
Stanford
Neurosciences
Program
.
As
an
undergraduate
at
the
University
of
Michigan
,
I
joined
the
Motor
Control
Laboratory
led
by
Dr
.
Samuel
Okonkwo
,
where
I
spent
three
years
bridging
experimental
and
computational
work
.
I
learned
to
perform
daily
signal
checks
on
neural
recording
hardware
,
preprocess
raw
voltage
traces
with
Kilosort
2.5,
and
curate
single
-
unit
clusters
by
hand
when
automated
algorithms
failed
on
noisy
channels
.
My
primary
project
involved
extending
a
standard
Kalman
filter
decoder
by
incorporating
recurrent
neural
network
priors
trained
on
historical
cursor
trajectories
.
This
hybrid
approach
improved
two
-
dimensional
reach
decoding
accuracy
from
seventy
-
one
percent
to
eighty
-
four
percent
in
closed
-
loop
sessions
and
reduced
mean
endpoint
error
by
nearly
two
millimeters
.
The
work
was
accepted
to
the
NeurIPS
2022
Workshop
on
Neuro
-
AI
as
a
contributed
talk
,
and
I
am
currently
a
co
-
first
author
on
a
journal
submission
that
formalizes
the
modeling
framework
and
validates
it
across
three
additional
non
-
human
primates
.
Beyond
the
specific
results
,
this
apprenticeship
taught
me
that
rigorous
experimental
practice
and
quantitative
theory
are
inseparable
:
a
beautiful
model
means
little
if
the
spike
sorting
is
sloppy
,
and
the
cleanest
dataset
cannot
compensate
for
a
decoder
that
ignores
motor
preparation
dynamics
.
After
graduating
,
I
moved
to
the
University
of
Cambridge
as
a
Gates
Cambridge
Scholar
to
work
with
Dr
.
Lena
Horváth
in
the
Cortical
Computation
Group
.
My
project
there
shifted
from
extracellular
electrophysiology
to
two
-
photon
calcium
imaging
in
mouse
primary
visual
cortex
during
an
orientation
-
discrimination
task
.
I
was
struck
by
how
strongly
behavioral
history
influenced
neural
responses
:
the
same
drifting
grating
could
elicit
markedly
different
population
activity
depending
on
whether
the
animal
had
previously
been
rewarded
for
a
similar
stimulus
.
To
capture
this
,
I
designed
a
transformer
-
based
model
that
incorporates
a
novel
history
-
attention
mechanism
,
allowing
the
network
to
flexibly
weight
past
trials
when
predicting
current
population
activity
.
The
model
outperformed
static
linear
decoders
by
a
significant
margin
and
revealed
that
history
-
dependent
effects
were
concentrated
in
superficial
layers
and
modulated
by
feedback
from
higher
visual
areas
.
I
also
learned
to
perform
cranial
window
implantations
,
run
intrinsic
optical
imaging
to
map
functional
regions
,
and
manage
a
data
pipeline
spanning
fifty
thousand
neurons
across
twelve
mice
.
A
preprint
describing
this
work
is
now
1 / 2
available
on
bioRxiv
,
and
we
are
preparing
a
submission
to
Nature
Neuroscience
that
situates
the
findings
within
the
broader
literature
on
serial
dependence
and
perceptual
decision
-
making
.
These
experiences
have
sharpened
my
central
research
question
:
how
does
the
brain
perform
credit
assignment
across
hierarchical
circuits
when
sensory
predictions
are
violated
?
I
am
convinced
that
answering
this
question
requires
a
two
-
way
dialogue
between
artificial
and
biological
neural
networks
.
On
one
side
,
normative
models
from
deep
learning
—
particularly
those
involving
feedback
alignment
,
target
propagation
,
and
meta
-
learning
—
offer
precise
,
testable
hypotheses
about
how
error
signals
could
be
routed
through
cortical
microcircuits
.
On
the
other
side
,
large
-
scale
electrophysiology
and
modern
circuit
-
tracing
methods
provide
the
empirical
ground
truth
needed
to
constrain
these
hypotheses
.
I
am
especially
eager
to
investigate
how
feedback
connections
,
often
treated
as
mere
modulatory
influences
,
might
carry
structured
teaching
signals
that
shape
feedforward
representational
geometry
during
perceptual
learning
.
This
line
of
inquiry
demands
a
program
that
values
both
quantitative
rigor
and
experimental
access
to
behaving
neural
circuits
.
Stanford
’
s
Neurosciences
PhD
Program
is
singularly
positioned
to
support
this
goal
.
The
opportunity
to
rotate
through
Dr
.
Krishna
Shenoy
’
s
laboratory
would
allow
me
to
deepen
my
expertise
in
dynamical
systems
approaches
to
motor
cortex
while
gaining
exposure
to
human
clinical
trials
of
neural
prostheses
.
Dr
.
Daniel
Yamins
’
work
on
goal
-
driven
deep
learning
models
of
the
ventral
visual
stream
—
particularly
the
CORnet
family
—
directly
informs
the
modeling
approaches
I
hope
to
extend
into
the
feedback
domain
.
Dr
.
Kalanit
Grill
-
Spector
’
s
investigations
of
human
visual
cortex
using
fMRI
and
electrocorticography
would
expose
me
to
methodologies
that
bridge
the
spatial
resolution
gap
between
single
-
unit
recordings
and
population
imaging
.
Beyond
individual
laboratories
,
the
Wu
Tsai
Neurosciences
Institute
and
the
existing
collaborations
between
the
Neurosciences
Program
,
Computer
Science
,
and
Bioengineering
embody
the
interdisciplinary
culture
that
I
believe
is
essential
for
modern
systems
neuroscience
.
I
am
also
drawn
to
the
program
’
s
emphasis
on
teaching
and
mentorship
,
values
I
have
tried
to
honor
by
mentoring
four
undergraduates
in
Cambridge
and
by
leading
workshops
on
Python
data
analysis
for
the
Michigan
Neuroscience
Graduate
Student
Organization
.
My
long
-
term
goal
is
to
lead
an
independent
research
laboratory
at
the
intersection
of
computational
neuroscience
and
neural
engineering
.
I
am
particularly
committed
to
developing
next
-
generation
neural
interfaces
that
restore
volitional
control
to
individuals
with
spinal
cord
injuries
,
a
mission
grounded
in
the
conviction
that
fundamental
insights
into
neural
computation
should
translate
into
tangible
clinical
benefit
.
Earning
a
PhD
at
Stanford
would
provide
the
methodological
breadth
,
theoretical
depth
,
and
collaborative
community
necessary
to
realize
this
vision
.
I
would
be
grateful
for
the
opportunity
to
contribute
to
and
learn
from
the
Stanford
neuroscience
community
,
and
I
thank
the
admissions
committee
for
considering
my
application
.
2 / 2