DAVID VON DOLLEN: Hello, my name is David Von Dollen.

And I'm here today to talk about layer wise learning

for quantum neural networks.

I'm an Area Lead and Technical Manager

for Volkswagen Group of America working on a variety of topics,

including quantum computing and machine learning out

of our Advanced Technologies Group

based here in San Francisco.

So to start, I'd like to give you

a little bit of an intuition about quantum neural networks.

A quantum neural network is a type

of circuit, where we have a register of qubits

with which we load some data, whether it's

classical or quantum.

For our project we looked at classical data, specifically

the MNIST data set.

And then you apply a series of unitary gates.

Now these can be random rotation gates, namely rotation y, x,

or z, and a series of Control Z gates.

Finally, we apply a readout on one qubit.

And with this we calculate a gradient for our parameters

for our unitaries.

Now a known problem for quantum neural networks

is what's called the barren plateau problem.

And essentially what it identifies

is that as the depth of a quantum neural network grows,

the variance of the gradients in randomly initialized quantum

neural networks decay exponentially

as a function of the number of qubits.

And so, given this problem, we developed this technique,

layerwise learning.

And so when we look at our technique,

we address the vanishing gradient problem.

But we also looked at using this new library, TensorFlow

Quantum, to train and experimentally verify

our algorithm.

The great thing about TensorFlow Quantum

is that it handles all of our training overhead.

And we can focus on research, rather than coding and getting

deep into the internals.

So looking at this vanishing gradient problem,

we may utilize larger gradients in shallow quantum neural

networks.

We can avoid configurations and random initialization,

which may lead to a barren plateau problem

when we apply layerwise learning.

And we can successively grow our quantum neural network layer

by layer by training, freezing, applying

another layer, freezing, and then also training and freezing

batches of layers.

So when we look at this layerwise learning,

we can think about this phase of sweeping

over the network, where we look at the first layer.

We train parameters.

We freeze.

We train our second parameter, and we freeze and so on.

In our second phase, we sweep through,

and we freeze batches of layers.

And when we do this, we find a speed up

in regards to training times.

And we also see a performance gain in our test error.

So when we looked at doing binary classification

for the digit 6 and 9 from MNIST,

we noticed an advantage when using 10 epochs per layer

in doing layerwise learning over what we call complete depth

learning, where we train all of the layers at once.

So to talk a little bit about TensorFlow Quantum,

we can generate our quantum neural network layers really

easily by using sympy and cirq to construct our circuit.

And then we can inject that using

a TFQ parametrized quantum circuit

layer into TensorFlow Keras.

And we can use the TensorFlow Keras loss functions

and optimizers to train the gradients for our parameters

for our quantum neural network.

And if you're interested in more detail,

we have an upcoming white paper.

This has been a really great collaboration

between Volkswagen and Google.

And if you have any questions, please

feel free to reach out to us.

Thank you very much.

[MUSIC PLAYING]