Transport, a process of converting quantum circuits into formats suitable for specific quantum hardware, presents a bottleneck that grows as quantum computers grow in complexity. Aleksander Kaczmarek of SoftServe Inc and Dikshant Dullal, Haique and their colleagues of Isaaq Pte Ltd are tackling this challenge by introducing methods to dramatically accelerate transparency times. Their research focuses on reuse of previously introduced circuits. This is a technique that avoids redundant calculations and significantly reduces computational costs, particularly when dealing with iterative processes such as layered learning in quantum machine learning. The team demonstrates that this approach, implemented within the rivet transformer pillar, achieves up to six times the speed improvements compared to traditional transport methods, paving the way for more efficient and scalable quantum algorithms.
Rivet Transpillar accelerates quantum machine learning
Quantum machine learning requires increasingly complex circuits, but preparing these circuits for real quantum hardware is a major challenge. This process, known as transport, transforms abstract quantum algorithms into a set of hardware-specific operations, often requiring substantial computational resources. This study introduces rivet transformer pillars, a new approach designed to accelerate this process and improve the efficiency of quantum machine learning applications. By reusing previously compiled circuit segments, rivets significantly reduce computational demands and accelerate the overall process. This innovation is especially valuable for algorithms such as layer-wise learning. Layerwise Learning allows circuits to be built in stages, allowing for efficient transport of additional layers.
The experiments show that the layered learning algorithm experiences an improvement of up to 600%, resulting in a significant reduction in transmission times. The researchers conducted extensive experiments using strategies to encode a variety of data, demonstrating consistent reductions in transport times in all ways. These strategies, such as angle encoding, ZzfeatureMap, and amplitude encoding, provide a unique trade-off between circuit complexity and expressivity, respectively. Rivets adapt to these differences, consistently providing improved performance, and ZzfeatureMap, which captures feature dependencies through intertwined gates, benefited in particular from rivet optimization. The results show that rivets not only reduce transpiration times, but also maintain accuracy and losses comparable to traditional training methods, making them a valuable tool for accelerating the research and development of quantum machine learning.
