Training variational GBS distributions

Technical details are available in the API documentation: sf.apps.train

Many quantum algorithms rely on the ability to train the parameters of quantum circuits, a strategy inspired by the success of neural networks in machine learning. Training is often performed by evaluating gradients of a cost function with respect to circuit parameters, then employing gradient-based optimization methods. In this demonstration, we outline the theoretical principles for training Gaussian Boson Sampling (GBS) circuits, first introduced in Ref. [1]. We then explain how to employ Strawberry Fields to perform the training by looking at basic examples in stochastic optimization and unsupervised learning. Let’s go! 🚀

Theory

As explained in more detail in Introduction to GBS, for a GBS device, the probability \(\Pr(S)\) of observing an output \(S=(s_1, s_2, \ldots, s_m)\), where \(s_i\) is the number of photons detected in the \(i\)-th mode, is given by

\[\Pr(S) = \frac{1}{\mathcal{N}} \frac{|\text{Haf}(A_S)|^2}{ s_1!\ldots s_m!},\]

where \(\mathcal{N}\) is a normalization constant and \(A\) is an arbitrary symmetric matrix with eigenvalues bounded between \(-1\) and \(1\). The matrix \(A\) can also be rescaled by a constant factor, which is equivalent to fixing a total mean photon number in the distribution.

We want to train this distribution to perform a specific task. For example, we may wish to reproduce the statistical properties of a given dataset, or to optimize the circuit to sample specific patterns with high probability. The strategy is to identify a parametrization of the distribution and compute gradients that allow us to optimize the parameters using gradient-based techniques. We refer to these circuits as variational GBS circuits, or VGBS for short.

Gradient formulas can be generally challenging to calculate, but there is a particular strategy known as the WAW parametrization (wow! what a great name) that leads to gradients that are simpler to compute. It involves transforming the symmetric matrix \(A\) as

\[A \rightarrow A_W = W A W,\]

where \(W = \text{diag}(\sqrt{w_1}, \sqrt{w_2}, \ldots, \sqrt{w_m})\) is a diagonal weight matrix and \(m\) is the number of modes. This parametrization is useful because the hafnian of \(A_W\) factorizes into two separate components

\[\text{Haf}(A_W) = \text{Haf}(A)\text{det}(W),\]

a property that can be cleverly exploited to compute gradients more efficiently. More broadly, it is convenient to embed trainable parameters \(\theta = (\theta_1, \ldots, \theta_d)\) into the weights \(w_k\). Several choices are possible, but here we focus on an exponential embedding

\[w_k = \exp(-\theta^T f^{(k)}),\]

where each \(f^{(k)}\) is a \(d\)-dimensional vector. The simplest case occurs by setting \(d=m\) and choosing these vectors to satisfy \(\theta^T f^{(k)} = \theta_k\) such that \(w_k = \exp(-\theta_k)\).

Training tasks

In stochastic optimization, we are given a function \(h(S)\) and the goal is optimize the parameters to sample from a distribution \(P_{\theta}(S)\) that minimizes the expectation value

\[C (\theta) = \sum_{S} h(S) P_{\theta}(S).\]

As shown in [1], the gradient of the cost function \(C (\theta)\) is given by

\[\partial_{\theta} C (\theta) = \sum_{S} h(S) P_{\theta}(S) \sum_{k=1}^{m} (s_k - \langle s_{k} \rangle) \partial_{\theta} \log w_{k},\]

where \(\langle s_k\rangle\) denotes the average photon number in mode \(k\). This gradient is an expectation value with respect to the GBS distribution, so it can be estimated by generating samples from the device. Convenient!

In a standard unsupervised learning scenario, data are assumed to be sampled from an unknown distribution and a common goal is to learn that distribution. More precisely, the goal is to use the data to train a model that can sample from a similar distribution, thus being able to generate new data. Training can be performed by minimizing the Kullback-Leibler (KL) divergence, which up to additive constants can be written as:

\[KL(\theta) = -\frac{1}{T}\sum_S \log[P_{\theta}(S)].\]

In this case \(S\) is an element of the data, \(P(S)\) is the probability of observing that element when sampling from the GBS distribution, and \(T\) is the total number of elements in the data. For the GBS distribution in the WAW parametrization, the gradient of the KL divergence can be written as

\[\partial_\theta KL(\theta) = - \sum_{k=1}^m\frac{1}{w_k}(\langle s_k\rangle_{\text{data}}- \langle s_k\rangle_{\text{GBS}})\partial_\theta w_k.\]

Remarkably, this gradient can be evaluated without a quantum computer because the mean photon numbers \(\langle s_k\rangle_{\text{GBS}}\) over the GBS distribution can be efficiently computed classically [1]. As we’ll show later, this leads to very fast training. This is true even if sampling the distribution remains classically intractable! 🤯

Stochastic optimization

We’re ready to start using Strawberry Fields to train GBS distributions. The main functions needed can be found in the train module, so let’s start by importing it.

from strawberryfields.apps import train

We explore a basic example where the goal is to optimize the distribution to favour photons appearing in a specific subset of modes, while minimizing the number of photons in the remaining modes. This can be achieved with the following cost function

import numpy as np

def h(s):
    not_subset = [k for k in range(len(s)) if k not in subset]
    return sum(s[not_subset]) - sum(s[subset])

The function is defined with respect to the subset of modes for which we want to observe many photons. This subset can be specified later on. Then, for a given sample s, we want to maximize the total number of photons in the subset. This can be achieved by minimizing its negative value, hence the term -sum(s[subset]). Similarly, for modes outside of the specified subset, we want to minimize their total sum, which explains the appearance of sum(s[not_subset]).

It’s now time to define the variational circuit. We’ll train a distribution based on a simple lollipop 🍭 graph with five nodes:

import networkx as nx
from strawberryfields.apps import plot

graph = nx.lollipop_graph(3, 2)
A = nx.to_numpy_array(graph)
plot.graph(graph)


Defining a variational GBS circuit consists of three steps: (i) specifying the embedding, (ii) building the circuit, and (iii) defining the cost function with respect to the circuit and embedding. We’ll go through each step one at a time. For the embedding of trainable parameters, we’ll use the simple form \(w_k = \exp(-\theta_k)\) outlined above, which can be accessed through the Exp class. Its only input is the number of modes in the device, which is equal to the number of nodes in the graph.

nr_modes = len(A)
weights = train.Exp(nr_modes)

Easy! The GBS distribution is determined by the symmetric matrix \(A\) — which we train using the WAW parametrization — and by the total mean photon number. There is freedom in choosing \(A\), but here we’ll just use the graph’s adjacency matrix. The total mean photon number is a hyperparameter of the distribution: in general, different choices may lead to different results in training. In fact, the mean photon number may change during training as a consequence of the weights being optimized. Finally, GBS devices can operate either with photon number-resolving detectors or threshold detectors, so there is an option to specify which one we intend to use. We’ll stick to detectors that can count photons.

n_mean = 6
vgbs = train.VGBS(A, n_mean, weights, threshold=False)

The last step before training is to define the cost function with respect to our previous choices. Since this is a stochastic optimization task, we employ the Stochastic class and input our previously defined cost function h.

During training, we’ll calculate gradients and evaluate the average of this cost function with respect to the GBS distribution. Both of these actions require estimating expectation values, so the number of samples in the estimation also needs to be specified. The parameters also need to be initialized. There is freedom in this choice, but here we’ll set them all to zero. Finally, we’ll aim to increase the number of photons in the “candy” part of the lollipop graph, which corresponds to the subset of modes [0, 1, 2].

np.random.seed(1969)  # for reproducibility
d = nr_modes
params = np.zeros(d)
subset = [0, 1, 2]
nr_samples = 100

print('Initial mean photon numbers = ', vgbs.mean_photons_by_mode(params))

Out:

Initial mean photon numbers =  [1.3273 1.3273 2.0721 0.9466 0.3268]

If training is successful, we should see the mean photon numbers of the first three modes increasing, while those of the last two modes become close to zero. We perform training over 200 steps of gradient descent with a learning rate of 0.01:

nr_steps = 200
rate = 0.01

for i in range(nr_steps):
    params -= rate * cost.grad(params, nr_samples)
    if i % 50 == 0:
        print('Cost = {:.3f}'.format(cost.evaluate(params, nr_samples)))

print('Final mean photon numbers = ', vgbs.mean_photons_by_mode(params))

Out:

Cost = -1.685
Cost = -3.892
Cost = -4.516
Cost = -4.677
Final mean photon numbers =  [3.1404 2.7654 1.1547 0.0931 0.0429]

Great! The cost function decreases smoothly and there is a clear increase in the mean photon numbers of the target modes, with a corresponding decrease in the remaining modes.

The transformed matrix \(A_W = W A W\) also needs to have eigenvalues bounded between -1 and 1, so continuing training indefinitely can lead to unphysical distributions when the weights become too large. It’s important to monitor this behaviour. We can confirm that the trained model is behaving according to plan by generating some samples. Although we still observe a couple of photons in the last two modes, most of the detections happen in the first three modes that we are targeting, just as intended.

Aw = vgbs.A(params)
samples = vgbs.generate_samples(Aw, n_samples=10)
print(samples)

Out:

[[0 0 0 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]
 [1 1 0 1 1]
 [0 0 0 0 0]
 [3 1 2 0 0]
 [2 2 2 0 0]
 [0 0 0 0 0]
 [0 0 0 0 0]
 [4 4 0 0 0]]

Unsupervised learning

We are going to train a circuit based on the pre-generated datasets in the data module. Gradients in this setting can be calculated efficiently, so we can study a larger graph with 30 nodes.

from strawberryfields.apps import data

data_pl = data.Planted()
data_samples = data_pl[:1000]  # we use only the first one thousand samples
A = data_pl.adj
nr_modes = len(A)

As before, we use the exponential embedding, but this time we define the VGBS circuit in threshold mode, since this is how the data samples were generated. The cost function is the Kullback-Liebler divergence, which depends on the data samples and can be accessed using KL:

weights = train.Exp(nr_modes)
n_mean = 1
vgbs = train.VGBS(A, n_mean, weights, threshold=True)
cost = train.KL(data_samples, vgbs)

We initialize parameters to zero and perform a longer optimization over one thousand steps with a learning rate of 0.15. This will allow us to reach a highly-trained model. Gradients can be computed efficiently, but evaluating the cost function is challenging because it requires calculating GBS probabilities, which generally take exponential time. Instead, we’ll keep track of the differences in mean photon numbers per mode for the data and model distributions.

from numpy.linalg import norm

params = np.zeros(nr_modes)
steps = 1000
rate = 0.15

for i in range(steps):
    params -= rate * cost.grad(params)
    if i % 100 == 0:
        diff = cost.mean_n_data - vgbs.mean_clicks_by_mode(params)
        print('Norm of difference = {:.5f}'.format(norm(diff)))

Out:

Norm of difference = 0.90856
Norm of difference = 0.01818
Norm of difference = 0.00532
Norm of difference = 0.00177
Norm of difference = 0.00062
Norm of difference = 0.00022
Norm of difference = 0.00008
Norm of difference = 0.00003
Norm of difference = 0.00001
Norm of difference = 0.00000

Wow! WAW! We reach almost perfect agreement between the data and the trained model. We can also generate a few samples, this time creating photon patterns that, in general, are not originally in the training data.

Aw = vgbs.A(params)
samples = vgbs.generate_samples(Aw, n_samples=10)
print(samples)

Out:

/opt/hostedtoolcache/Python/3.7.17/x64/lib/python3.7/site-packages/thewalrus/samples.py:367: ComplexWarning:

Casting complex values to real discards the imaginary part

/opt/hostedtoolcache/Python/3.7.17/x64/lib/python3.7/site-packages/thewalrus/samples.py:370: ComplexWarning:

Casting complex values to real discards the imaginary part

[[0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 1]
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [1 1 1 1 1 0 1 1 1 1 1 0 1 0 1 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 0 1 0]
 [0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 1 0 0 0 0 0 1 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 1 0 0 0 0 0 0 0 0 0 0 0 1 0 0 1 1 0 0 0 0 0 1 0 0 0 0 1]
 [0 0 1 1 0 0 1 0 0 1 0 0 0 1 0 1 1 0 1 1 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]]

The examples we have covered are introductory tasks aimed at mastering the basics of training variational GBS circuits. These ideas are new and there is much left to explore in terms of their scope and extensions. For example, the original paper [1] studies how GBS devices can be trained to find solutions to max clique problems. What new applications come to your mind?

References

1(1,2,3,4)

Leonardo Banchi, Nicolás Quesada, and Juan Miguel Arrazola. Training Gaussian Boson Sampling Distributions. arXiv:2004.04770. 2020.

Total running time of the script: ( 0 minutes 15.445 seconds)

Gallery generated by Sphinx-Gallery