Speaker
Description
Diffusion Models (DMs) are a leading class of physics-inspired generative AI (GenAI) models. DMs obtained state-of-the-art results with Denoising Diffusion Probabilistic Models (DDPMs), which progressively add noise to data and then reverse the process by learning the noise added at each forward time step (denoising). Despite their benefits, they require intensive computational resources, long training procedures and many diffusion steps to permit the model to learn how to map the features from pure noisy samples to original ones. In parallel, due to the rise of quantum computing and its potential to represent exponentially large systems with a polynomial number
of qubits, there has been a growing interest in quantum GenAI that has yielded to the introduction of quantum diffusion models (QDMs). The majority of the works in the context of QDMs propose to implement the backward part with a parametrized quantum circuit that is iteratively trained to minimize an appropriate loss function.
Our formulation goes in the opposite direction: we execute the forward dynamics on a real noisy intermediate-scale quantum (NISQ) device of IBM to take advantage of the intrinsic quantum noise of the machine, and use a classical learning model for data generation. We are able to generate 2D and 3D medical data of moderate and large dimensions, validating our results by computing distinct state-of-the-art metrics usually used in the domain of GenAI such as Frechét Inception Distance (FID), Structural Similarity Index Measure (SSIM), and we compare the obtained results with a classical discrete state-space DM.
| Topic | Hybrid quantum-classical approaches for learning and optimization |
|---|