A mathematical theory of relational generalization in transitive inference
Proceedings of the National Academy of Sciences,
Год журнала:
2024,
Номер
121(28)
Опубликована: Июль 5, 2024
Humans
and
animals
routinely
infer
relations
between
different
items
or
events
generalize
these
to
novel
combinations
of
items.
This
allows
them
respond
appropriately
radically
circumstances
is
fundamental
advanced
cognition.
However,
how
learning
systems
(including
the
brain)
can
implement
necessary
inductive
biases
has
been
unclear.
We
investigated
transitive
inference
(TI),
a
classic
relational
task
paradigm
in
which
subjects
must
learn
relation
(
A
>
B
C
)
it
new
).
Through
mathematical
analysis,
we
found
that
broad
range
biologically
relevant
models
(e.g.
gradient
flow
ridge
regression)
perform
TI
successfully
recapitulate
signature
behavioral
patterns
long
observed
living
subjects.
First,
with
item-wise
additive
representations
automatically
encode
relations.
Second,
for
more
general
representations,
single
scalar
“conjunctivity
factor”
determines
model
behavior
on
and,
further,
principle
norm
minimization
(a
standard
statistical
bias)
enables
fixed,
partly
conjunctive
transitively.
Finally,
neural
networks
“rich
regime,”
representation
improves
generalization
many
tasks,
unexpectedly
show
poor
anomalous
TI.
find
such
form
(over
hidden
weights)
yields
local
encoding
mechanism
lacking
transitivity.
Our
findings
minimal
principles
give
rise
classical
bias
(transitivity),
explain
empirically
behaviors,
establish
formal
approach
understanding
basis
abstraction.
Язык: Английский
Neural mechanisms of relational learning and fast knowledge reassembly in plastic neural networks
Nature Neuroscience,
Год журнала:
2025,
Номер
unknown
Опубликована: Янв. 15, 2025
Язык: Английский
An active neural mechanism for relational learning and fast knowledge reassembly
bioRxiv (Cold Spring Harbor Laboratory),
Год журнала:
2023,
Номер
unknown
Опубликована: Июль 28, 2023
How
do
we
gain
general
insights
from
limited
novel
experiences?
Humans
and
animals
have
a
striking
ability
to
learn
relationships
between
experienced
items,
enabling
efficient
generalization
rapid
assimilation
of
new
information.
One
fundamental
instance
such
relational
learning
is
transitive
inference
(learn
Язык: Английский
Neural dynamics of robust legged robots
Frontiers in Robotics and AI,
Год журнала:
2024,
Номер
11
Опубликована: Апрель 18, 2024
Legged
robot
control
has
improved
in
recent
years
with
the
rise
of
deep
reinforcement
learning,
however,
much
underlying
neural
mechanisms
remain
difficult
to
interpret.
Our
aim
is
leverage
bio-inspired
methods
from
computational
neuroscience
better
understand
activity
robust
locomotion
controllers.
Similar
past
work,
we
observe
that
terrain-based
curriculum
learning
improves
agent
stability.
We
study
biomechanical
responses
and
within
our
network
controller
by
simultaneously
pairing
physical
disturbances
targeted
ablations.
identify
an
agile
hip
reflex
enables
regain
its
balance
recover
lateral
perturbations.
Model
gradients
are
employed
quantify
relative
degree
various
sensory
feedback
channels
drive
this
reflexive
behavior.
also
find
recurrent
dynamics
implicated
behavior,
utilize
sampling-based
ablation
these
key
neurons.
framework
combines
model-based
for
drawing
causal
relationships
between
embodied
Язык: Английский
Transitive inference as probabilistic preference learning
Psychonomic Bulletin & Review,
Год журнала:
2024,
Номер
unknown
Опубликована: Окт. 22, 2024
Язык: Английский
Learning to infer transitively: serial ordering on a mental line in premotor cortex
bioRxiv (Cold Spring Harbor Laboratory),
Год журнала:
2024,
Номер
unknown
Опубликована: Окт. 29, 2024
ABSTRACT
Transitive
inference
(TI)
is
a
form
of
deductive
reasoning
that
allows
to
infer
unknown
relationships
among
premises.
It
hypothesized
this
cognitive
task
accomplished
by
mapping
stimuli
onto
linear
workspace,
referred
as
the
‘mental
line,’
based
on
their
arbitrarily
assigned
ranks.
However,
open
questions
remain:
does
mental
line
have
neural
correlate,
and
if
so,
where
how
it
represented
learned
in
brain?
In
study,
we
investigate
role
monkeys’
dorsal
premotor
cortex
(PMd)
encoding
during
acquisition
item
relationships.
Our
findings
provide
evidence
TI
can
be
solved
through
transformation
representations
ranked
items.
We
show
PMd
multi-unit
activity
organizes
along
theoretically
informed
direction,
implementing
geometrical
solution
effectively
explains
animal
behavior.
results
suggest
plays
crucial
integrating
into
‘geometric
symbolic
distance
(i.e.,
rank
difference)
between
items
influences
related
motor
decisions.
Furthermore,
observe
an
ongoing
learning
process
characterized
rotation
line,
which
aligns
manifold
plan
unfolds.
This
elucidates
cortical
optimization
strategy
statistical
structure
task.
Язык: Английский
Prefrontal cortex contribution in transitive inference task through the interplay of beta and gamma oscillations
Communications Biology,
Год журнала:
2024,
Номер
7(1)
Опубликована: Дек. 31, 2024
Transitive
inference
allows
people
to
infer
new
relations
between
previously
experienced
premises.
It
has
been
hypothesized
that
this
logical
thinking
relies
on
a
mental
schema
spatially
organizes
elements,
facilitating
inferential
insights.
However,
recent
evidence
challenges
the
need
for
these
complex
cognitive
processes.
To
dig
into
neural
substrate
driving
TI
processes,
we
examine
role
of
beta
and
gamma
local
field
potential
bands
in
prefrontal
cortex
2
monkeys.
During
problem-solving
period,
discover
tight
link
modulation
complexity.
This
correlation
diminishes
its
strength
before
initiating
motor
response,
indicating
chosen
item.
Notably,
while
band
maintains
constant
relationship
with
performance
throughout
trial,
shows
flexible
relationship.
research
highlights
interplay
computations
when
solving
problems.
Язык: Английский