Bottleneck layers achieve advantages through matrix product operators and unleash circuits

The challenge of implementing complex artificial intelligence with Quantum Computers takes a step forward with all of Román Orús' research and colleagues from Borja Aizpurua, Sukhbinder Singh and Multiverse Computing. They explore how to translate key components of existing classical neural networks into quantum hardware, and aim to unlock the potential speed advantages of new quantum devices. Teams accomplish this by compressing complex layers into a more manageable format and skillfully splitting the processing between quantum computers and traditional hardware. Validated with standard image classification tasks such as MNIST and CIFAR-10, this hybrid approach presents a promising pathway for integrating quantum computation with established machine learning techniques, and could ultimately accelerate the development of more powerful artificial intelligence.

The team will use dispersers of tensor networks to explore new approaches, providing a promising way to use quantum resources more efficiently to represent and process classical data. This work is attributed to growing interest in dispersive quantum algorithms. This utilizes quantum computers as co-processors to optimize quantum circuit parameters to minimize cost functions. An important obstacle is the quantum bottleneck. This is how difficult it is to load classical data into quantum states into quantum states with no overhead in terms of quantum gates and circuit depth.

The purpose of this work is to mitigate this bottleneck by investigating data encoding strategies that utilize the structure of tensor networks to reduce the required quantum resources. Specifically, this study focuses on image classification, demonstrating that this approach can be used to effectively implement classical neural networks in quantum devices, achieving performance comparable to classical counterparts, while utilizing fewer quantum resources. The team represents the weights and activation of a classical convolutional neural network as a tensor network, using a disassembly procedure to map this network to quantum circuits, establishing a pathway for deploying classical machine learning models in short-term quantum devices. Removing the bottleneck layer from classically trained neural networks on quantum computers represents the pathway to achieving quantum advantage in short-term devices.

This approach starts with compression and represents the target linear layer as an effective matrix product performance (MPO) without decomposition of the model's performance. This MPO is then unleashed into an even more compact form, allowing for a hybrid classic quartile execution scheme. Here, the unleashed MPO runs on classic hardware and unravels the circuitry to the Quantum computer. The team introduces two complementary algorithms for MPO solution growth, designed to optimize processes and maximize efficiency.

Quantum tensor networks compress classic models

This study explores the intersection of quantum computing and machine learning, focusing on how quantum technology can enhance or compress classic machine learning models. The use of tensor networks inspired by quantum entanglement and quantum circuitry, and the emphasis on representing and manipulating data more efficiently, especially in large-scale models. Tensor networks provide a way to represent high-dimensional data in a more compact and efficient way, filling the gap between quantum concepts and classical machine learning, compressing large-scale models, efficiently representing data, and accelerating computation. This study simulates quantum computing in classical computers using tensor networks, allowing researchers to explore the potential benefits of quantum machine learning without requiring access to actual quantum hardware.

The team investigates using tensor networks to compress large language models (LLMs) to reduce size and computational costs while maintaining performance. Data sets such as MNIST and CIFAR-10 are used to evaluate the performance of machine learning algorithms. This study recognizes the challenges in scaling tensor networks and quantum algorithms for processing very large data sets and models, and investigates techniques such as positive bias and unlabeled contraction to improve efficiency. Hybrid quantum classic approaches may be important to maximize the potential of these technologies. The team investigates specific algorithms and methods, including excessive approximate contractions, sine-free tensor contractions, distracted angled representations, and quantum circuit-born machines. Quantum computing and tensor networks provide promising means to improve machine learning, but there remain major challenges in terms of scalability and algorithm development.

Quantum Bottleneck Layer for Hybrid Computing

This study demonstrates how to convert classical neural network components into a hybrid classical quantum computing framework. The team compressed and converted the bottle net layers from pre-trained networks originally used for image classification tasks into quantum circuits, allowing these circuits to work with classical hardware. This approach aims to reduce the computational burden of classical processors by offloading specific tasks to quantum computers without sacrificing the overall accuracy of neural networks, highlighting the importance of efficient circuit design to achieve practical quantum advantages. This method has successfully freed complex weight matrices into more manageable quantum circuits, but further investigation of circuit structure and gate selection is required to minimize complexity. Future work could focus on optimizing circuit geometry, investigating the placement of specific gates, and investigating how to reduce the computational costs and tomography steps of measurements within quantum workflows. Combining multiple layers into a single circuit can result in more efficient benefits, and reconstructing measurements as computational resources rather than cost.