Toward Next-Generation Artificial Intelligence: Catalyzing the NeuroAI Revolution — part 2/3

Cleaned-text code points [13327, 28625). Bibliography omitted here; source.md includes it.

crates
falling
off
trucks
leading
to
bad
outcomes.
And
even
if
it
has
been
trained
on
the
dangers
of
falling
crates,
the
system
might
consider
an
empty
plastic
bag
being
blown
out
of
the
car
in
front
of
it
as
an
obstacle
to
avoid
at
all
costs
rather
than
an
irritant,
again,
because
it
doesn’t
actually
understand
what
a
plastic
bag
is
or
how
unthreatening
it
is
physically.
This
inability
to
handle
scenarios
that
have
not
appeared in the training data is a significant challenge to widespread reliance on AI systems.
To
be
successful
in
an
unpredictable
and
changing
world,
an
agent
must
be
flexible
and
master
novel
situations
by
using
its
general
knowledge
about
how
such
situations
are
likely
to
unfold.
This
is
arguably
what
animals
do.
Animals
are
born
with
most
of
the
skills
needed
to
thrive
or
can
rapidly
acquire
them
from
limited
experience,
thanks
to
their
strong
foundation
in
real-world
interaction,
courtesy
of
evolution
and
development
36
.
Thus,
it
is
clear
that
training
from
scratch
for
a
specific
task
is
not
how
animals
obtain
their
impressive
skills;
animals
do
not
arrive
into
the
world
tabula
rasa
and
then
rely
on
large
labeled
training
sets
to
learn.
Although
machine
learning
has
been
pursuing
approaches
for
sidestepping
this
tabula
rasa
limitation,
including
self-supervised
learning,
transfer
learning,
continual
learning,
meta
learning,
one-shot
learning
and
imitation
learning
37
,
none
of
these
approaches
comes
close
to
achieving
the
flexibility
found
in
most
animals.
Thus,
we
argue
that
understanding
the
neural
circuit-level
principles
that
provide
the
foundation
for
behavioral
flexibility
in
the
real-world,
even
in
simple
animals,
has
the
potential
to
greatly
increase
the
flexibility
and
utility
of
AI
systems.
Put
another
way,
we
can
greatly
accelerate
our
search
for
general-purpose
circuits
for
real-world
interaction
by
taking
advantage
of
the
optimization process that evolution has already engaged in
38–45
.
Animals
compute
efficiently.
One
important
challenge
for
modern
AI
–
that
our
brains
have
overcome
–
is
energy
efficiency.
Training
a
neural
network
requires
enormous
amounts
of
energy.
For
example,
training
a
large
language
model
such
as
GPT-3
requires
over
1000
megawatts-hours,
enough
to
power
a
small
town
for
a
day
46
.
Biological
systems
are,
by
contrast,
much
more
energy
efficient:
The
human
brain
uses
about
20
watts
47
.
The
difference
in
energy
requirement
between
brains
and
computers
derives
from
differences
in
information
processing.
First,
at
an
algorithmic
level,
modern
large-scale
ANNs,
such
as
large
language
models
26
,
rely
on
very
large
feedforward
architectures
with
self-attention
to
process
sequences
over
time
23
,
ignoring
the
potential
power
of
recurrence
for
processing
sequential
information.
One
reason
for
this
is
that
currently
we
do
not
have
efficient
mechanisms
for
credit
assignment
calculations
in
recurrent
networks.
In
contrast,
brains
utilize
flexible
recurrent
architectures
that
can
solve
the
temporal
credit
assignment
problem
with
great
efficiency.
Uncovering
the
mechanisms
by
which
this
happens
could
potentially
enable
us
to
increase
the
energy
efficiency
of
artificial
systems.
Alternatively,
it
has
been
proposed
that
the
synaptic
dynamics
within
adjacent
dendritic
spines
could
serve
as
a
mechanism
for
learning
sequential
structure,
a
scheme
that
could
potentially
be
efficiently
implemented
in
hardware
48
.
Second,
at
an
implementation
level,
neural
circuits
differ
from
digital
computers.
Neural
circuits
compute
effectively
despite
the
presence
of
unreliable
or
"noisy"
components.
For
example,
synaptic
release,
the
primary
means
of
communication
between
neurons,
can
be
so
unreliable
that
only
one
in
every
ten
messages
is
transmitted
49
.
Furthermore,
neurons
interact
mainly
by
transmitting
action
potentials
(spikes),
an
asynchronous
communication
protocol.
Like
the
interactions
between
conventional
digital
elements,
the
output
of
a
neuron
can
be
viewed
as
a
string
of
0s
and
1s;
but
unlike
a
digital
computer,
the
energy
cost
of
a
“1”
(i.e.
of
a
spike)
is
several
orders
of
magnitude
higher
than
that
of
a
“0”
50
.
Because
biological
circuits
operate
in
a
regime
where
spikes
are
sparse
–
even
very
active
neurons

rarely
fire
at
more
than
100
spikes
per
second
and
typical
cortical
firing
rates
may
be
less
than
1
spike/second
–
they
are
much
more
energy
efficient
51
.
Spike-based
computation
has
also
been
shown
to
be
orders of magnitude faster and more energy efficient in recent hardware implementation
52
.
A roadmap for solving the embodied Turing test
How
might
artificial
systems
that
pass
the
embodied
Turing
test
be
developed?
One
natural
approach
would
be
to
do
so
incrementally,
guided
by
our
evolutionary
history.
For
example,
almost
all
animals
engage
in
goal-directed
locomotion;
they
move
toward
some
stimuli
(e.g.
food
sources)
and
away
from
others
(e.g.
threats).
Layered
on
top
of
these
foundational
abilities
are
more
sophisticated
skills,
such
the
ability
to
combine
different
streams
of
sensory
information
(e.g.
visual
and
olfactory),
to
use
this
sensory
information
to
distinguish
food
sources
and
threats,
to
navigate
to
previous
locations,
to
weigh
possible
rewards
and
threats
to
achieve
goals,
and
to
interact
with
the
world
in
precise
ways
in
service
of
these
goals.
Most
of
these
–
and
many
other
–
sophisticated
abilities
are
found
to
some
extent
in
even
very
simple
organisms,
such
as
worms.
In
more
complex
animals,
such
as
fish
and
mammals,
these
abilities
are
elaborated and combined with new strategies to enable more powerful behavioral strategies.
This
evolutionary
perspective
suggests
a
strategy
for
passing
the
embodied
Turing
test
by
breaking
it
down
into
a
series
of
incrementally
challenging
ones
that
build
on
each
other,
and
iteratively
optimizing
on
this
series
53
.
Specifically,
the
embodied
Turing
test
comprises
challenges
that
include
a
wide
range
of
organisms
used
in
neuroscience
research,
including
worms,
flies,
fish,
rodents
and
primates.
This
would
enable
us
to
deploy
the
vast
amount
of
knowledge
we
have
accumulated
about
the
behavior,
biomechanics,
and
neural
mechanisms
of
these
model
organisms
to
both
precisely
define
each
species-specific
embodied
Turing
test
and
serve
as
strong
inductive
biases
to
guide
the
development
of
robust AI controllers that can pass it.
The
performance
of
these
artificial
agents
could
be
compared
with
that
of
animals.
Rich
behavioral
datasets
representing
a
large
swath
of
a
species’
ethogram
have
now
been
collected
and
can
be
deployed
to
benchmark
performance
on
species-specific
embodied
Turing
tests.
Furthermore,
these
datasets
are
being
rapidly
expanded
given
new
tools
in
3D
videography
54–57
.
Additionally,
detailed
biomechanical
measurements
support
highly
realistic
animal
body
models,
complete
with
skeletal
constraints,
muscles,
tendons,
and
paw
features
58
.
Combined
with
the
open-sourcing
of
powerful,
fast
physics
simulators
and
virtual
environments
59,60
,
these
models
will
afford
the
opportunity
for
embodied
Turing
test
research
to
be
performed
in
silico
at
scale
33
.
Finally,
existing
extensive
neural
datasets
with
simultaneous
neural
recordings
across
multiple
brain
regions
during
behavior,
combined
with
increasingly
detailed
neural
anatomy
and
connectomics,
provide
a
powerful
roadmap
for
the
design
of
AI
systems
that
can
control
virtual
animals
to
recapitulate
the
behaviors
of
their
in
vivo
counterparts
and
thus
pass
the
embodied
Turing test.
Importantly,
the
specifics
of
the
embodied
Turing
test
for
each
species
can
be
tuned
to
the
needs
of
different
groups
of
researchers.
We
can
test
the
capacity
of
AI
systems
in
terms
of
sensorimotor
control,
self-supervised
and
continual
learning,
generalization,
memory-guided
behavior
on
both
short
and

life-long
timescales,
and
social
interactions.
Despite
these
potentially
different
areas
of
interest,
the
challenges
that
compose
the
embodied
Turing
test
can
be
standardized
to
permit
the
quantification
of
progress
and
comparison
between
research
efforts.
Standardization
can
be
fostered
by
stakeholders
including
government
and
private
funders,
large
research
organizations
such
as
the
Allen
Institute,
and
major
collaborations
like
the
International
Brain
Lab,
with
an
eye
toward
the
development
of
common
APIs
and
support
for
competitions
as
has
been
an
important
impetus
for
much
progress
in
machine
learning
and
robotics
61,62
.
Ultimately,
virtual
organisms
that
demonstrate
successful
recapitulation
of
behaviors
of
interest
can
be
adapted
to
the
physical
world
with
additional
efforts
in
robotics
and
deployed
to solve real-world problems.
What we need
Achieving
these
goals
will
require
significant
resources
deployed
in
three
main
areas.
First,
we
must
train
a
new
generation
of
AI
researchers
who
are
equally
at
home
in
engineering/computational
science
and
neuroscience
.
These
researchers
will
chart
fundamentally
new
directions
in
AI
research
by
drawing
on
decades
of
progress
in
neuroscience.
The
greatest
challenge
will
be
in
determining
how
to
exploit
the
synergies
and
overlaps
in
neuroscience,
computational
science,
and
other
relevant
fields
to
advance
our
quest:
identifying
what
details
of
the
brain’s
circuitry,
biophysics,
and
chemistry
are
important
and
what
can
be
disregarded
in
the
application
to
AI.
There
is
thus
a
critical
need
for
researchers
with
dual
training
in
AI
and
neuroscience
to
apply
insights
from
neuroscience
to
advance
AI
and
to
help
design
experiments
that
generate
new
insights
relevant
to
AI.
Although
there
is
already
some
research
of
this
type,
it
exists
largely
at
the
margins
of
mainstream
neuroscience;
training
in
neuroscience
has
thus
far
been
motivated
and
funded
mainly
by
the
goal
of
improving
human
health
and
of
understanding
the
brain
as
such.
This
lack
of
alignment
between
fields
might
explain,
e.g.
,
the
multi-decade
gap
between
Hubel
and
Wiesel’s
discovery
of
the
structure
of
the
visual
system
6
and
the
development
and
application
of
convolutional
neural
networks
in
modern
machine
learning
8
.
Thus,
the
success
of
a
NeuroAI
research
program
depends
on
the
formation
of
a
community
of
researchers
for
whom
the
raison
d'être
of
their
training
is
to
exploit
synergies
between
neuroscience
and
AI.
Explicit
design
of
new
training
programs
can
ensure
that
the
NeuroAI
research
community
reflects
the
demographics
of
society
as
a
whole
and
is
equipped
with
the
ethical tools needed to ensure that the development of AI benefits society
63
.
Second,
we
must
create
a
shared
platform
capable
of
developing
and
testing
these
virtual
agents.
One
of
the
greatest
technical
challenges
that
we
will
face
in
creating
an
iterative,
embodied
Turing
test
and
evolving
artificial
organisms
to
pass
it
is
the
amount
of
computational
power
required.
Currently,
training
just
one
large
neural
network
model
on
a
single
embodied
task
(e.g.
control
of
a
body
in
3-dimensional
space)
can
take
days
on
specialized
distributed
hardware
64
.
For
multiple
research
groups
to
iteratively
work
together
to
optimize
and
evaluate
a
large
number
of
agents
over
multiple
generations
on
increasingly
complex
embodied
Turing
tests,
a
large
investment
in
a
shared
computational
platform
will
be
required.
Much
like
a
particle
accelerator
in
physics
or
large
telescope
in
astronomy,
this
sort
of
large-scale
shared
resource
will
be
essential
for
moving
the
brain-inspired
AI
research
agenda
forward.
It
will
require
a
major
organizational
effort,
with
government
and
ideally
also
industry
support,
that
has
as
its central goal scientific progress on animal and human-like intelligence.

Third,
we
must
support
fundamental
theoretical
and
experimental
research
on
neural
computation
.
We
have
learned
a
tremendous
amount
about
the
brain
over
the
last
decades,
through
the
efforts
of
the
NIH,
in
no
small
measure
due
to
the
BRAIN
Initiative,
and
other
major
funders
,
and
we
are
now
reaching
an
understanding
of
the
vast
diversity
of
the
brain’s
individual
cellular
elements,
neurons,
and
how
they
function
as
parts
of
simple
circuits.
With
these
building
blocks
in
place,
we
are
poised
to
shift
our
focus
toward
understanding
how
the
brain
functions
as
an
integrated
intelligent
system.
This
will
require
insight
into
how
a
hundred
billion
neurons
of
a
thousand
different
types,
each
one
communicating
with
thousands
of
other
neurons,
with
variable,
adaptable
connections,
are
wired
together,
and
the
computational
capabilities
–
the
intelligence
–
that
emerges.
We
must
reverse
engineer
the
brain
to
abstract
the
underlying
principles.
Taking
advantage
of
the
powerful
synergies
between
neuroscience
and
AI
will
require
program
and
infrastructure
support
to
organize
and
enable
research
across
the
disciplines
at
a
large
scale.
Fortunately,
there
is
now
broad
political
agreement
that
investments
in
AI
research
are
essential
to
humanity’s
technological
future.
Indeed,
IARPA
(Intelligence
Advanced
Research
Projects
Activity)
was
a
pioneer
in
this
field,
launching
the
Machine
Intelligence
from
Cortical
Networks
(MICrONS)
project.
This
project
spearheaded
the
collection
of
an
unprecedented
data
set
consisting
of
a
portion
of
a
mouse
connectome
and
associated
functional
responses
with
the
specific
goal
of
catalyzing
the
development
of
next-generation
AI
algorithms
65
.
Nonetheless,
community-wide
efforts
to
bridge
the
fields
of
neuroscience
and
AI
will
require
robust
investments
from
government
resources,
as
well
as
oversight
of
project
milestones,
commercialization
support,
ethics,
and
big
bets
on
innovative
ideas.
In
the
U.S.,
there
are
currently
some
lines
of
federal
resourcing,
such
as
the
NSF’s
National
Artificial
Intelligence
Research
Institutes,
explicitly
dedicated
to
driving
innovation
and
discovery
in
AI
from
neuroscience
research,
but
these
are
largely
designed
to
support
a
traditional
academic
model
with
different
groups
investigating
different
questions,
rather
than
the
creation
of
a
centralized
effort
that
could
create
something
like
the
embodied
Turing
test.
Likewise,
AI
support
grants
in
the
U.S.
are
predominantly
ancillary
programs
through
the
NIH,
NSF,
DoD,
and
even
the
EPA
–
each
of
which
have
their
own
directives
and
goals
–
and
this
pattern
is
shared
by
funding
agencies
globally.
This
leaves
a
significant
funding
gap
for
technology
development
as
an
end
in
itself.
The
creation
o
f
overarching
directives
either
through
existing
entities,
or
as
a
stand-alone
agency,
to
support
NeuroAI
and
AI
research
would
drive
this
mission
and
consequently
unlock the potential for AI to benefit humanity.
Conclusions
Despite
the
long
history
of
neuroscience
driving
advances
in
AI
and
the
tremendous
potential
for
future
advances,
most
engineers
and
computational
scientists
in
the
field
are
unaware
of
the
history
and
opportunities.
The
influence
of
neuroscience
on
shaping
the
thinking
of
von
Neumann,
Turing
and
other
giants
of
computational
theory
are
rarely
mentioned
in
a
typical
computer
science
curriculum.
Leading
AI
conferences
such
as
NeurIPS,
which
once
served
to
showcase
the
latest
advances
in
both
computational
neuroscience
and
machine
learning,
now
focus
almost
exclusively
on
the
latter.
Even
some
researchers
aware
of
the
historical
importance
of
neuroscience
in
shaping
the
field
often
argue
that
it
has
lost
its