Quantum machine learning and simulation: A new technology area

Quantum generative adversarial networks (qGAN)

With increasing computation demands, ways for AI-based applications are intensely investigated, which combine GANs with quantum-based algorithmics, which then get summarized as (Quantum generative adversarial networks). These networks can transform (e.g., recorded scenes, such as converting daytime scenes to nighttime) without requiring additional recordings.

By applying this technology to automotive and avionic data, the team demonstrated its practical and impactful capabilities, even with today’s reduced quantum hardware availability.

Our experiment was implemented on AWS Braket, applying hybrid jobs. Consequently, it can be ported to any other platform that provides gate-based QPU, as well as classical computing power (e.g., IBM Qiskit or simulation environments built on top, or GPUs with CUDA-Q libraries).

Implementation details of qGANs

The core of our approach involves a classical generative adversarial network (GAN) enhanced with quantum elements in a hybrid-model manner.

[INSERT GRAPHIC]

The solution obeys the following general design properties:

• Classical neural network: We use a classical neural network for the forward and backward generators. This network is responsible for transforming images, such as converting daytime scenes into nighttime scenes.

• Quantum discriminator: We integrate a quantum discriminator to improve the training process. This component compares the generated images with the original ones to optimize the transformation.

An abstract model structure is shown in Figure 2, visualizing the three core elements as background-colored boxes and the next level of contents.

Results from the Quantum discriminator are crucial for calculating loss functions (defined in this case as a measure of the identity of two pictures), as shown below. The identity of pictures is measured first as a distance between the original picture X and a picture that went through forward G and backward F generators and is supposed to be original again (in an ideal world).

[INSERT GRAPHIC]

And measured second as a distance between the transformed picture Y and a picture  that went first through backward F, then forward G generators. 

[INSERT GRAPHIC]

During our experiments, we tested two options for a Quantum discriminator using two different quantum circuits. Our circuits had four steps:

1. Amplitude encoding of an image

2. QFT for image entropy calculation or Controlled SWAP for quantum distance

3. Measurement

4. Decoding results

A schematic visualization of quantum circuits can be seen in Figure 4.

[INSERT GRAPHIC]

Figure 4a presents the Quantum Fourier Transformation to analyze the image’s entropy. Figure 4b is a Quantum K-Means distance, which was used to compare two pictures.

Findings

The following section provides initial insights into the results of the qGAN experiments applied to a use case inspired by ABQCC. One obvious question is around indications of the number of qubit dependencies in GAN describing metrics.

[INSERT GRAPHIC]

Our experiments demonstrate that the loss function converges significantly within 200 epochs, approaching zero as the number of iterations increases. We tested the model with 4, 8 and 16 qubits, observing that increasing the number of qubits leads to faster and better convergence of the loss function. Further experimental results suggest a log(n) dependency, but it requires more experiments on bigger quantum computers or simulators.

Performance comparison

For understanding the applicability and, ultimately, a business value estimation, a basic understanding can be generated by understanding timing, thus computing the times required for executing such domain conversions. This is summarized as follows:

[INSERT TABLE 1]

[INSERT TABLE 2]

The above tables represent two types of performance comparison: E2E, which includes data encoding and decoding, as well as context switching between the classical and quantum parts, and isolated measurement of the quantum task. All times are for one picture of the training set.

• Simulators: We compared the end-to-end timing for processing one image across different simulators. The local state vector simulator is the fastest but is limited by local computing power, handling up to around 20 qubits.

• Quantum processing units (QPUs): We also compared the isolated duration of quantum tasks on various backends, including cloud-based state vector simulators and real QPUs like IQM Garnet and Rigetti Ankaa-3. The cloud-based simulator and Rigetti Ankaa-3 showed comparable execution times, while IQM Garnet was slower.

Crucially, we accessed real quantum backends (IQM and Rigetti) via AWS Braket services (not directly). That might have had an impact on the presented results relating to E2E execution time, because classical computing power is located within AWS data centers, and quantum computing is remote.

During the execution of our algorithms, we require access to decent computing power (mainly GPU). The training algorithm is a hybrid, comprising a significant portion of classical machine learning. Our hybrid jobs used several types of EC2 instances (ml.m5.large, ml.p3.2xlarge, ml.p3.16xlarge, ml.p4d.24xlarge).

While the advantages of quantum computing are, from a theoretical perspective, clear for certain mathematical problems, DXC also acknowledges the challenges arising from hardware availability.

Currently, real quantum hardware is limited in availability, and the uptime of some QPUs may not yet be sufficient for processing sophisticated datasets to support productive usage. Additionally, encoding and decoding data into quantum states can sometimes negate the computational advantages.

Summing up

DXC's approach to quantum computing emphasizes the ease of experimentation with classical, quantum and hybrid systems. Utilizing platforms like AWS Braket (and others, such as NVIDIA CUDA-Q and IBM Qiskit), DXC experiments demonstrated computational speedups by reducing complexity.

It’s worth mentioning that during ABQCC, we used and detailed the implementation of qGANs (via integrating quantum computing with classical ML algorithm parts), which showed substantial advancements in the performance of isolated discriminator step.

The scalability of these solutions with the growth of QPU power provides further prospects for the potential of such hybrid solutions, even though the availability of QPUs varies. Some offer 24/7 access, whereas others have more limited uptimes. Therefore, a current and concrete algorithm implementation is still closely tied to the hardware and software layers.

Challenges and risk mitigation associated with the use of emerging technologies can be addressed through experienced and multidisciplinary teams provided by DXC.

Despite these challenges, the promise of QML and simulation is undeniable. DXC's innovative solutions are paving the way for a future where quantum computing plays a central role in solving some of the most complex problems across industries.

As this technology continues to evolve, the potential for groundbreaking advancements grows, making it an exciting field to work in and lead.