AI offers boundless potential,
from enhancing health care diagnosis to making
our cities smarter, empowering us with education. The potential is immense. While AI can
revolutionize so much, realizing this potential
also rests on our ability to solve some really important
technical and societal challenges. So I would like to talk with
you about a new idea for machine learning we termed
liquid networks. We began to develop this
work as a way of addressing some of the challenges that we
have with today's AI solutions. Because despite the
great opportunities, we also have plenty of
technical challenges that remain to be solved. So first among the AI
challenges is the data itself. We require huge
amounts of data that gets fed in immense models. And these models have huge
computational and environmental costs. We also have an issue
with data quality because if the data
quality is not high, the performance of the
model will not be good. Bad data means bad performance. Furthermore, we have
these black box systems where it's really
impossible to find out how the system makes decisions. And this is really
problematic, especially for safety critical
applications. So let me show you the
essential idea behind liquid AI. And I will show this to you in
the context of an autonomous driving application. And then we can
generalize to others. So here is a self-driving car,
which was built by our students at MIT using traditional
deep neural networks. And it does pretty well. It was trained in the city. It drives really well in
a completely different environment. It can make decisions
at intersections. It can recognize the goal. So it's pretty good, right? But let me open the
hood to show you how this vehicle makes decisions. So you will see in the
right-hand corner the map. You will see in the upper left
corner the camera input stream. And the decision-making engine
is the big rectangular box in the middle with blue
and yellow blinking lights. And there are about
100,000 artificial neurons that are working together
to tell this car what to do. And it is absolutely
impossible to correlate how the neurons
activate with what the vehicle does because
there are too many of them. There's also half a
million parameters. Take a look at the
lower left-hand corner, where we see the attention map. This is where, in the image,
the vehicle looks in order to make decisions. You see how noisy it is? You see how this vehicle is
looking at the bushes and at the trees on the
side of the road? So this is a bit of a problem. Well, I would like to do better. I would like a vehicle whose
decisions I can understand. And in fact, with
liquid networks, we have a new class of models. And here, you can see the
liquid network solution for the same problem. Now you will see
the entire model consisting of 19 artificial
neurons, liquid neurons. And look at the attention map. Look how clean it
is and how focused it is on the road horizon and
on the sides of the road, which is how I drive. And so liquid networks seem to
understand their task better than deep networks. And because they
are so compact, they have many other properties. So in particular, we can
take the output of 19 neurons and turn them into
a decision tree. Now, that could show the humans
how these networks decide. And so they are much
closer to a world where we can have
machine learning that is understandable. We can apply liquid networks
to many other applications. Here is a solution
consisting of 11 neurons. And this is driving a plane in
a canyon of unknown geometry. The plane has to hit these
points at unknown locations. And it's really extraordinary
that all you need is 11 artificial neurons,
liquid network neurons, in order to solve this problem. So how did we accomplish this? Well, we started by the
continuous time neural network framework. And in continuous networks,
the solution, or the neuron, is defined by a series of
differential equations. And these models are kind of
temporal neural network, which includes standard recurrent
neural networks, also neural ODEs, continuous
time (CT) RNNs, and now, liquid networks. And it's really
extraordinary because, by using differential equations
and by using continuous time networks, we can model very
elegantly complex problems, like problems that
involve physical dynamics. For instance, in
this case, we have the half-cheetah standards. And it can be modeled elegantly
with these continuous time networks. However, when you take an
existing continuous time solution and you model
even a simple problem, like can you get this
half cheetah to walk, you actually get performance
that is not that much better than a standard LSTM. And so, however, with liquid
networks, you can do better. OK, so how do we achieve
this better performance? Well, we achieve it with two
mathematical innovations. First of all, we
change the equation that defines the
activity of the neuron. We start with a linear
state space model, and then we introduce
non-linearities over the synaptic connections. And then when we plug these
two equations into each other, we end up with this equation. And so what's interesting
about this equation is that the time constant that
should go in front of x of t is actually dependent on x of t. And this allows us to have
neural network solutions that are able to change their
underlying equations based on the input that they
see after training. We also do some other changes,
like we change the wiring architecture of the network. And you can read about
this in our papers. And so, now, let's go
back to the attention of a whole suite of networks,
CNNs, (CT) RNNs, LSTMs, and other solutions. So back to the driving
in lane problem, you'll see that all previous
solutions are really looking at the context,
not at the actual task. And in fact, we have
a mathematical basis for this result. We can actually
prove that our liquid network solutions are causal. In other words, they
connect cause and effect in ways that are consistent with
the mathematical definitions of causality. Now, I promised you
a fast solution. But these networks are defined
by differential equations. So you might ask, do they really
need numerical problem solvers, because that would actually
be a huge computational hit. Well, it turns out
we have a closed form solution for the hairy equation
that goes inside the neuron. And the solution
has a good bound. It's good enough. And you can see in this chart,
in red, the ODE solution, and in blue, the solution
with our approximation. And you see that they are really
quite close to each other. So these liquid networks can
learn causal relationships because they form causal models. Unlike other models defined
by differential equations like neural ODEs and
(CT) RNNs, in essence, these networks recognize when
their outputs are being changed by certain interventions. And then they learn how to
correlate cause and effect. All right, so let me
give you a final example to convince you that these
networks are really valuable. So here, we have a
different problem. We are training a drone
how to fly in the woods. Notice that it's summertime. So we give our drones
examples of videos like you'll see in this example. And these are not
annotated in any way. And we train a
variety of models, for instance, a standard
deep neural network. And now, when we get the
standard network trained in that environment to find
the object and go to it, you see that the model
has a lot of trouble. The attention is very noisy. And then also notice that
the background is different because now it's fall time. So the context of
the task has changed. Because deep networks are
so dependent on context, they don't do so well. But look at our liquid
network solution. They are so focused on the task. And the drone has no
problem finding the object. We can further go all
the way to the winter with the same model
trained in the summer. And we get a good solution. And finally, we can even
change the context of the task entirely. We can put it in an
urban environment. And we can go from a static
object to a dynamic object. The same model trained in
the summer in the woods does well in this example. So this is, again,
because we have a provably causal solution. So liquid networks are a new
model for machine learning. They are compact,
interpretable, and causal. And they have shown great
promise in generalization under heavy distribution shifts. Thank you. [APPLAUSE]
