Quantum Diffusion Models

Abstract: We propose a quantum version of a generative diffusion model. In this algorithm, artificial neural networks are replaced with parameterized quantum circuits, in order to directly generate quantum states. We present both a full quantum and a latent quantum version of the algorithm; we also present a conditioned version of these models. The models' performances have been evaluated using quantitative metrics complemented by qualitative assessments. An implementation of a simplified version of the algorithm has been executed on real NISQ quantum hardware.

• The paper introduces innovative quantum diffusion models that use parameterized quantum circuits to generate quantum states directly.
• It proposes a full quantum model alongside a latent quantum-classical hybrid variant, incorporating bottleneck architectures and infidelity loss for training.
• Simulation and IBM hardware experiments reveal improved sample quality and parameter reduction compared to classical diffusion models.

Quantum Diffusion Models

This essay analyzes the manuscript "Quantum Diffusion Models" (2311.15444), which proposes innovative quantum adaptations of classical diffusion models, leveraging parameterized quantum circuits (PQC) to facilitate the direct generation of quantum states. The paper introduces a full quantum model and a latent quantum-classical hybrid model, presenting results of implementations on both simulations and actual NISQ hardware. The findings suggest potential enhancements over classical counterparts in terms of parameter efficiency and capabilities for generating quantum data.

Background and Methodology

Parameterized Quantum Circuits

At the foundation of the proposed quantum diffusion models is the use of PQCs, which are analogous to artificial neural networks (ANNs) but operate within the quantum domain. A typical PQC comprises layers of rotation gates with trainable parameters, interspersed with C-NOT gates to induce qubit entanglement.

Figure 1: Example of a PQC layer ansatz for three qubits. The rotation gates contain classical, trainable parameters while the C-NOT gates create entanglement between qubits.

The PQCs in this study were trained using classical optimization techniques relying on simulation environments due to the noise constraints of current quantum hardware. This constraint necessitates an amplitude encoding scheme to translate classical data into quantum states, allowing for exponential feature encoding.

Classical Diffusion Models

Diffusion models traditionally focus on mapping an arbitrary distribution to a tractable Gaussian distribution through a forward Markov chain, followed by the training of an ANN to approximate the reverse process. The adoption of parameterized quantum circuits modifies this paradigm due to direct quantum state manipulation, circumventing intermediate classical processes that are computationally expensive on quantum hardware.

Quantum Diffusion Model

Model Architecture

The quantum diffusion model replaces classical neural network components with PQCs. The quantum model's architecture is divided into a full quantum model designed for directly generating quantum states and a latent model that employs a classical autoencoder.

Figure 2: Representation of the bottleneck (left) and reverse-bottleneck architectures (right) using $m=1$. Each unitary transformation block employs a layered structure.

To mitigate the limitations imposed by existing quantum hardware, particularly regarding qubit decoherence and connectivity, innovative PQC architectures are introduced. The "bottleneck" and "reverse-bottleneck" configurations significantly influence performance, where intermediate measurements and ancillary qubit interactions replace conventional layer transformations.

Model Training

Training in a quantum context involves the parameterized transformation of qubit states. Here, infidelity loss is employed as the optimization criterion—a choice that aligns with quantum computing methodologies by utilizing fidelity tests.

Simulation and Hardware Results

Simulation Insights

Simulations on MNIST data, downscaled for quantum compatibility, reveal the potential of quantum diffusion models to achieve outcomes with fewer parameters, observing improved sample quality in latent spaces compared to classical models. For instance, the use of label qubits in conditioned models significantly enhances discriminative sampling.

Figure 3: Samples generated from the full quantum model.

Hardware Implementation

The implementation on IBM's quantum device, ibm_hanoi, addresses connectivity and error constraints through strategic reductions in PQC complexity, facilitated by a reduced number of denoising steps and modified gate arrangements.

Figure 4: Architecture of IBM_hanoi quantum computer. The C-NOT connectivity and T1 error probabilities are highlighted.

Empirical results illustrate the enduring challenge of noise, particularly its decoherence effects, but also demonstrate the hardware's capacity to approximate classical results accurately under proper configurations.

Conclusion

The "Quantum Diffusion Models" paper lays the groundwork for future advancements in quantum generative modeling, particularly highlighting the opportunities in parameter minimization and scalability through quantum state representation. The integration of quantum methods provides a path for potential breakthroughs in processing power and complexity, harboring implications for data domains unreachable by classical algorithms. As quantum hardware progresses, innovations in noise mitigation and full quantum training regimes are anticipated to drive further developments in this area.