today I’m here to tell you about what we
can do with the brain of a little worm C
elegans is a one millimetre long world
with a rather simple brain it can
exhibits remarkable behavioral
plasticity that the current state
state-of-the-art AI systems are yet to
produce given the capacity its nervous
system contains only three hundred and
two cells that are hardwired by
approximately around eight thousand
neurons and with that simple brain it
can still learn it it can mate
it can complex it can process complex
chemicals and complex sensory data and
also it can adopt search mechanisms in
the environment C elegans is amongst the
world’s best understood animals its
entire nervous system was mapped in 1986
and it provided a substrate for research
and asked the questions of how the brain
gives rise to behavior so i malay rai
scientists i’m fascinated by how much c
elegans can get with such little
computational resources especially given
the capacity and given the AI systems
that you’re developing these days then
when I look like closer to the
architecture of the brain so there are
certain features that I want to like
discusses with you guys so the first
thing is that I see a particular
hierarchical architecture of the nervous
system that has that is actually
specific to C elegans and I’ll tell you
more about it then the network is not an
all-too all connected network so every
node is connect is not connected to all
part of this nervous system it there are
certain neuronal pathways so the network
is a sparse and we know that synaptic
transmission in nervous systems in
general are much more complex than the
way that we’re currently treating AI
systems and this is also to read how
neuronal dynamics are realized and in
actual nervous systems so I want to have
these kind of features and ask myself
can I have an AI system that can produce
such similar behavior and be as
expressive as the elegance given those
features with together with our teams at
Vienna University of Technology at
Institute of Science and Technology in
Institute of molecular pathology in
Vienna and MIT we create these systems
so we create systems that are similar
and closer to the closer to the
capacities of natural learning systems
so I would like to also especially thank
and acknowledge my colleagues
contributions to this work that we are
working together closely at all levels
concept design and implementation we
designed this system so we got those
four features of the nervous system and
we designed and we formalize them in in
a form of an artificial and new
artificial intelligence system that
incorporates the hierarchical design so
to process input data and to generate an
output similar to what C elegans has and
also we incorporated some rules that
adapt the sparsity of the network within
this platform we also got closer in
mathematical expressions than that
synaptic propagation have as well we
actually have a richer synaptic
propagation and richard synaptic
connections so we for formulated
actually a kind of dynamical system and
now we want to solve
AI tasks with such systems so here I’m
showing you a control dashboard so on
this dashboard on the right side you see
one example circuit small circuit
inspired by the brain
the worm that wants to control an
inverted pendulum problem so actually
the environment on the Left I’m sure so
basically the circuit wants to control
and balance the pole on an upright
position and on the rock on the top side
I’m showing a learning curve during the
learning process that the network was
getting tuned in order to maximize the
amount of reward in the system so that
we can collect how much we can keep this
pendulum actually upward so let’s see so
right now the red dot on the learning
curve shows that where we are at the
training so at the beginning of the
process you see the circuit cannot
realize the behavior so the pendulum
Falls constantly but then the more the
system is getting learned we see that we
are trying to maximize actually the
capacity of this system and now for
example now we are in the middle of the
curve Steele the circuit is trying to
capture the behavior now I should work
better because you’re higher and we
learned much more in this system and
then once we reach on the right as part
of the learning curve the system has
learned how to control this pendulum
like this scenario let me show you
another example this is actually another
control problem where we have an
underpowered car that we want to create
a momentum so this circuit is trying to
create a momentum for the car to drive
it uphill and now I’m showing you one
successful episode of training in this
system so as we see we are the circuit
is interacting and interplay with each
other in order to create that and now we
have that so now let’s see another
episode of that where you can look at
the circuit and dynamics of this system
that is trying to create such momentum
and now so we also extended it in
real-life applications
okay that was simulated platforms and
then we extended it to autonomous kind
of parking of of some robots in our lab
and as we see the task is done with the
same same kind of circuit that I showed
you before so what’s the next step that
we designed these platforms and that now
what we do we try to quantify and we try
to understand the behavior of these
circuits that we have for example here
I’m showing the neurons that that are
involved in the in controlling those
tasks and I’m projecting the activity of
one node to the output in order to see
what form of dynamics this network is
creating and with some quantitative
methods that we developed we want to
understand what happens inside the
systems so after and then this actually
increases also the expressivity of the
network so what what why do we have to
do this like you have artificial neural
networks that already can handle these
problems so what are the advantages
wider why you’re doing this so we
actually try to compare this to other
artificial intelligence systems that are
available and I want to count three
advantages of this system the first one
is in environment such as full
autonomous parking scenarios so actually
at vu here okay so this car is actually
moving in the environment trying to find
a parking spot and then performing a
part in trajectory and also it can avoid
collisions so with with obstacles on
this way then in those kind of tasks
then with the simplest artificial neural
network we could do for this specific
task beach 943 trainable parameters we
could solve this problem and then
with our systems we could do that with
only 49 tradable parameters so our
networks are smaller in size than
current artificial intelligence systems
we know that environment is always comes
with uncertainty and there are so many
interaction and disturbances on the
world like noise like for example an
autonomous car that is moving in the
street so all of a sudden environment
can cause rain and you know sensors of
the robot can get distracted so noise is
the parameters that always attack a
learning system so then if you try to
have a curve where we are increasing on
the horizontal axis we are increasing
the amount of noise of the environment
and we measure the output disturbances
on the system we want to compare this
between artificial intelligence methods
and our methods and this is how our
system is reacting so these systems are
highly resilient to noise so as as we
compared to their equivalent artificial
intelligence in tasks such as autonomous
parking and the third advantage that I
want to count is the most important one
these systems are more interpretable
they are expressive we can understand
then an internal their internal dynamics
so on the left side I am showing you a
simple artificial neural network that
handle these formal problems and as we
see this is such a complex kind of
architecture in order to understand its
behavior so on the right I’m showing one
of our networks that actually can solve
the similar problem with much less
parameters and also like having it’s
more expressive kind of neuronal
dynamics so it networks with such
advantages we are getting closer to have
control systems that are safer such as
autonomous cars that you’re working on
right now and we know that autonomous
cars and environments with
high uncertainty are gonna come are
gonna soon become the everyday norm for
artificial intelligence agents thank you [Applause]
can do with the brain of a little worm C
elegans is a one millimetre long world
with a rather simple brain it can
exhibits remarkable behavioral
plasticity that the current state
state-of-the-art AI systems are yet to
produce given the capacity its nervous
system contains only three hundred and
two cells that are hardwired by
approximately around eight thousand
neurons and with that simple brain it
can still learn it it can mate
it can complex it can process complex
chemicals and complex sensory data and
also it can adopt search mechanisms in
the environment C elegans is amongst the
world’s best understood animals its
entire nervous system was mapped in 1986
and it provided a substrate for research
and asked the questions of how the brain
gives rise to behavior so i malay rai
scientists i’m fascinated by how much c
elegans can get with such little
computational resources especially given
the capacity and given the AI systems
that you’re developing these days then
when I look like closer to the
architecture of the brain so there are
certain features that I want to like
discusses with you guys so the first
thing is that I see a particular
hierarchical architecture of the nervous
system that has that is actually
specific to C elegans and I’ll tell you
more about it then the network is not an
all-too all connected network so every
node is connect is not connected to all
part of this nervous system it there are
certain neuronal pathways so the network
is a sparse and we know that synaptic
transmission in nervous systems in
general are much more complex than the
way that we’re currently treating AI
systems and this is also to read how
neuronal dynamics are realized and in
actual nervous systems so I want to have
these kind of features and ask myself
can I have an AI system that can produce
such similar behavior and be as
expressive as the elegance given those
features with together with our teams at
Vienna University of Technology at
Institute of Science and Technology in
Institute of molecular pathology in
Vienna and MIT we create these systems
so we create systems that are similar
and closer to the closer to the
capacities of natural learning systems
so I would like to also especially thank
and acknowledge my colleagues
contributions to this work that we are
working together closely at all levels
concept design and implementation we
designed this system so we got those
four features of the nervous system and
we designed and we formalize them in in
a form of an artificial and new
artificial intelligence system that
incorporates the hierarchical design so
to process input data and to generate an
output similar to what C elegans has and
also we incorporated some rules that
adapt the sparsity of the network within
this platform we also got closer in
mathematical expressions than that
synaptic propagation have as well we
actually have a richer synaptic
propagation and richard synaptic
connections so we for formulated
actually a kind of dynamical system and
now we want to solve
AI tasks with such systems so here I’m
showing you a control dashboard so on
this dashboard on the right side you see
one example circuit small circuit
inspired by the brain
the worm that wants to control an
inverted pendulum problem so actually
the environment on the Left I’m sure so
basically the circuit wants to control
and balance the pole on an upright
position and on the rock on the top side
I’m showing a learning curve during the
learning process that the network was
getting tuned in order to maximize the
amount of reward in the system so that
we can collect how much we can keep this
pendulum actually upward so let’s see so
right now the red dot on the learning
curve shows that where we are at the
training so at the beginning of the
process you see the circuit cannot
realize the behavior so the pendulum
Falls constantly but then the more the
system is getting learned we see that we
are trying to maximize actually the
capacity of this system and now for
example now we are in the middle of the
curve Steele the circuit is trying to
capture the behavior now I should work
better because you’re higher and we
learned much more in this system and
then once we reach on the right as part
of the learning curve the system has
learned how to control this pendulum
like this scenario let me show you
another example this is actually another
control problem where we have an
underpowered car that we want to create
a momentum so this circuit is trying to
create a momentum for the car to drive
it uphill and now I’m showing you one
successful episode of training in this
system so as we see we are the circuit
is interacting and interplay with each
other in order to create that and now we
have that so now let’s see another
episode of that where you can look at
the circuit and dynamics of this system
that is trying to create such momentum
and now so we also extended it in
real-life applications
okay that was simulated platforms and
then we extended it to autonomous kind
of parking of of some robots in our lab
and as we see the task is done with the
same same kind of circuit that I showed
you before so what’s the next step that
we designed these platforms and that now
what we do we try to quantify and we try
to understand the behavior of these
circuits that we have for example here
I’m showing the neurons that that are
involved in the in controlling those
tasks and I’m projecting the activity of
one node to the output in order to see
what form of dynamics this network is
creating and with some quantitative
methods that we developed we want to
understand what happens inside the
systems so after and then this actually
increases also the expressivity of the
network so what what why do we have to
do this like you have artificial neural
networks that already can handle these
problems so what are the advantages
wider why you’re doing this so we
actually try to compare this to other
artificial intelligence systems that are
available and I want to count three
advantages of this system the first one
is in environment such as full
autonomous parking scenarios so actually
at vu here okay so this car is actually
moving in the environment trying to find
a parking spot and then performing a
part in trajectory and also it can avoid
collisions so with with obstacles on
this way then in those kind of tasks
then with the simplest artificial neural
network we could do for this specific
task beach 943 trainable parameters we
could solve this problem and then
with our systems we could do that with
only 49 tradable parameters so our
networks are smaller in size than
current artificial intelligence systems
we know that environment is always comes
with uncertainty and there are so many
interaction and disturbances on the
world like noise like for example an
autonomous car that is moving in the
street so all of a sudden environment
can cause rain and you know sensors of
the robot can get distracted so noise is
the parameters that always attack a
learning system so then if you try to
have a curve where we are increasing on
the horizontal axis we are increasing
the amount of noise of the environment
and we measure the output disturbances
on the system we want to compare this
between artificial intelligence methods
and our methods and this is how our
system is reacting so these systems are
highly resilient to noise so as as we
compared to their equivalent artificial
intelligence in tasks such as autonomous
parking and the third advantage that I
want to count is the most important one
these systems are more interpretable
they are expressive we can understand
then an internal their internal dynamics
so on the left side I am showing you a
simple artificial neural network that
handle these formal problems and as we
see this is such a complex kind of
architecture in order to understand its
behavior so on the right I’m showing one
of our networks that actually can solve
the similar problem with much less
parameters and also like having it’s
more expressive kind of neuronal
dynamics so it networks with such
advantages we are getting closer to have
control systems that are safer such as
autonomous cars that you’re working on
right now and we know that autonomous
cars and environments with
high uncertainty are gonna come are
gonna soon become the everyday norm for
artificial intelligence agents thank you [Applause]
Please follow and like us: