The Week in Quantum Computing - October 21st - QuEra investment from Google, Quandela, v-Score, Quantum for LLMs
The Week in Quantum Computing - October 21st - QuEra investment from Google, Quandela, v-Score, Quantum for LLMs
Issue #206
Quick Recap
Quandela, a French startup, outlined its ambitious roadmap to achieve 50 logical qubits by 2028, leveraging its photonic quantum computing technology. QuEra Computing secured a strategic investment from Google Quantum AI to enhance its neutral atom technology, focusing on quantum error correction. In the UK, Nu Quantum introduced a qubit-photon interface to connect multiple quantum processors, addressing the scaling challenges of quantum computing. In Europe, IQM Quantum Computers was selected to deliver advanced quantum systems as part of the Euro-Q-Exa hybrid system, enhancing Europe's high-performance computing infrastructure. Additionally, new research from Harvard and Google Quantum AI demonstrated the potential of quantum memory to exponentially enhance computational power, suggesting that quantum advantage could be achieved sooner than expected. A new paper suggest a path for quantum computers to enhance LLMs, or at least make them less energy hungry.
Is that so? Quantum Parameter Adaptation for parameter efficient learning?
I.e. recuce the number of parameters required to train LLMs
This approach aims to reduce the number of parameters required for fine-tuning large language models like GPT-2 and Gemma-2, without sacrificing performance. For Quantum Parameter Generation, the approach leverages Quantum Neural Networks (QNNs) to generate classical model weights during the training process, thus decoupling inference from quantum hardware. QPA then combines QNNs with classical multi-layer perceptron mapping models to generate parameters for fine-tuning methods like Low-Rank Adaptation.
Experimental Results: The method demonstrated significant parameter reductions (52.06% for GPT-2 and 16.84% for Gemma-2) while maintaining or slightly improving performance in text generation tasks.The research aligns with and builds upon existing works in parameter-efficient fine-tuning and QML. Previous studies have explored PEFT methods such as LoRA, DoRA, and Prefix-Tuning, aiming to reduce trainable parameters in LLMs.
QML research has shown theoretical benefits but faced practical challenges like data encoding and reliance on quantum hardware during inference. The proposed QPA method addresses these challenges by maintaining classical hardware for inference and using quantum methods only during training.
HoweveR:
- Theoretically, QPA presents a promising approach to reducing parameters in LLMs, but its transformative potential depends on practical implementation and scalability. Yeah, SCALABILITY. The point many tend to ignore.
- Challenges in quantum hardware availability and the practicalities of running large-scale QNNs could hinder widespread adoption.
- While the decoupling of quantum resources during inference is beneficial, the necessity of quantum hardware during training might limit its application in production environments.
- The scalability of QPA for models beyond the sizes tested (e.g., models with tens or hundreds of billions of parameters) remains uncertain. Nobody would use a GPT2 today for any production workload. No matter how cheap the model is.
- The current results are promising for up to 2 billion parameters, but further validation is needed for larger models and more complex tasks.
- The computational overhead and memory requirements for simulating quantum circuits on classical hardware might pose significant challenges for large-scale production use.
*Results have the same probability as quantum advantage happening in 2025.
The Week in Quantum Computing
Quandela wants 50 logical qubits by 2028
Quandela, a French quantum computing startup, aims to achieve 50 logical qubits by 2028, with its first logical qubit targeted for 2025. The company plans to increase quantum operations per second (QOPS) from 400 to 10,000 and open a second factory by 2027. Co-founder and CEO Niccolo Somaschi describes their roadmap as "ambitious and credible." Founded in 2017, Quandela specializes in photonic quantum computing, using photons for information processing, which offers scalability and stability advantages. The company is part of the PROQCIMA program under France 2030, aiming to develop universal quantum computers by 2032. Quandela's focus on photonics positions it uniquely in the evolving quantum landscape, promising significant advancements by 2030.
https://www.quandela.com/technology/roadmap/
QuEra Computing announces investment from key strategic partner to accelerate development of large-scale, fault-tolerant quantum computers
QuEra Computing has secured a strategic investment from Google Quantum AI to advance its development of large-scale, fault-tolerant quantum computers using neutral atom technology. This collaboration builds on research from Harvard and MIT, led by Mikhail Lukin, Vladan Vuletic, and Markus Greiner. The investment aims to enhance QuEra's quantum error correction capabilities and further its strategic roadmap outlined in January 2024. Andy Ory, Interim CEO of QuEra, stated, "This investment...positions us as the recognized market leader for neutral atom-based quantum computing solutions." QuEra's technology targets industries like finance, pharmaceuticals, and energy, potentially enabling novel AI and machine learning applications.
New Interface Uses Light to Scale Up Quantum Computers
Nu Quantum, a UK-based startup, has developed a qubit-photon interface to connect multiple quantum processors, potentially scaling quantum computers into larger, more powerful machines. This innovation aims to address the challenge of scaling quantum computers, which currently operate with noisy intermediate-scale quantum platforms of hundreds of qubits. Carmen Palacios-Berraquero, Nu Quantum's CEO, emphasizes the need for modular, interconnected quantum processing units rather than larger, inefficient machines. The device uses a microscopic cavity to enhance qubit-photon interactions, stabilizing cavity lengths to 80 picometers.
https://spectrum.ieee.org/quantum-network-interface
A Blueprint for Canadian Deep Tech Leadership from Quantum Industry Canada
Canada has emerged as a global leader in quantum technology, driven by early strategic investments in research and commercialization. Lisa Lambert, CEO of Quantum Industry Canada (QIC), highlights Canada's pioneering role in the second quantum revolution, which involves creating and manipulating quantum states for practical applications. Canada is home to the first commercial quantum computing company, D-Wave, and the first quantum software company, 1QBit. The country boasts the second highest number of quantum startups globally, with notable achievements from companies like Xanadu and Nord Quantique. QIC, a national consortium, aims to leverage these strengths for global business success.
Quantum computing and photonics discovery potentially shrinks critical parts by 1,000 times
Researchers have made a discovery that could make quantum computing more compact, potentially shrinking essential components 1,000 times while also requiring less equipment. The research is published in Nature Photonics. Nanyang Technological University, Singapore (NTU Singapore) scientists have found a way to address this approach's problem by producing linked pairs of photons using much thinner materials that are just 1.2 micrometers thick, or about 80 times thinner than a strand of hair. And they did so without needing additional optical gear to maintain the link between the photon pairs, making the overall set-up simpler. "Our novel method to create entangled photon pairs paves the way for making quantum optical entanglement sources much smaller, which will be critical for applications in quantum information and photonic quantum computing," said NTU's Prof Gao Weibo, who led the researchers.
https://phys.org/news/2024-10-quantum-photonics-discovery-potentially-critical.html
IQM Selected to Deliver Two Advanced Quantum Computers as Part of Euro-Q-Exa Hybrid System
IQM Quantum Computers has been selected by the EuroHPC Joint Undertaking to deliver two advanced Radiance quantum computers, a 54-qubit system in 2025 and a 150-qubit system in 2026, as part of the Euro-Q-Exa hybrid system. These systems will be integrated into the Leibniz Supercomputing Centre (LRZ) in Germany, enhancing Europe's high-performance computing infrastructure. The initiative is funded by EuroHPC JU, the German Federal Ministry of Education and Research, and the Bavarian State Ministry of Sciences and the Arts. Markus Blume, Bavarian State Minister, emphasized the strategic importance of this collaboration, stating, "54 qubits doesn't sound like much, but it is the gateway to a whole new universe."
Quantum Memory Proves Exponentially Powerful
Recent advancements in quantum computing reveal that quantum memory can exponentially enhance computational power. Sitan Chen from Harvard University demonstrated that using just two copies of a quantum state can significantly reduce the need for repeated measurements, a finding echoed by a team at Google Quantum AI. Richard Kueng from Johannes Kepler University Linz emphasized the unexpected power of these multi-copy measurements. This breakthrough suggests that quantum memory could establish quantum advantage by requiring less data rather than fewer computational steps. Hsin-Yuan Huang from Google Quantum AI noted this could achieve quantum advantage sooner. As Jarrod McClean from Google Quantum AI highlighted, these developments bring us closer to practical applications in understanding complex quantum systems.
https://www.quantamagazine.org/quantum-memory-proves-exponentially-powerful-20241016/
Hybrid quantum error correction technique integrates continuous and discrete variables
A groundbreaking approach in quantum computing involves a hybrid quantum error correction technique that combines continuous variables (CV) and discrete variables (DV) qubits. This innovative architecture, developed by researchers, employs hybrid fusion techniques to interconnect these hybrid qubits, forming an error-correcting lattice structure. This method aims to enhance fault-tolerant quantum computing by leveraging the strengths of both CV and DV qubits. The integration of these qubit types could potentially lead to more robust and scalable quantum systems.
https://phys.org/news/2024-10-hybrid-quantum-error-technique-discrete.html
UNM Engineering professor part of $7.5M DOE quantum computing research
Milad Marvian, an assistant professor at the University of New Mexico (UNM) School of Engineering, is participating in a $7.5 million Department of Energy (DOE) project aimed at advancing quantum algorithm theory. Marvian, a member of the Quantum New Mexico Institute, is collaborating on the "FAR-Qu: Fundamental Algorithmic Research towards Quantum Utility" project, led by Sandia National Laboratories. This initiative involves partnerships with Caltech, University of Maryland, and several national laboratories. The project is part of a $65 million DOE investment across 10 projects to enhance quantum computing research in the U.S. Marvian aims to "develop novel mathematical tools" to deepen understanding of quantum computation.
https://news.unm.edu/news/unm-engineering-professor-part-of-7-5m-doe-quantum-computing-research
National Tsing Hua University Unveils World's Smallest Room-Temperature Quantum Computer Using Single Photon.
Researchers at National Tsing Hua University, led by Professor Chuu Chih-sung, have developed the world's smallest quantum computer using a single photon. This breakthrough, published in *Physical Review Applied*, addresses key challenges in quantum computing, such as high energy costs and low-temperature requirements. The team encoded information in 32 time bins of a photon, allowing stable operation at room temperature. University President John Kao highlighted the compactness compared to large cooling systems in other labs. The National Science and Technology Council is hosting Nobel laureate Alain Aspect at Quantum Taiwan to foster global collaboration.
https://www.taipeitimes.com/News/front/archives/2024/10/17/2003825430
Title:A Quantum Circuit-Based Compression Perspective for Parameter-Efficient Learning
Chen-Yu Liu, Chao-Han Huck Yang, Min-Hsiu Hsieh, and Hsi-Sheng Goan introduce Quantum Parameter Adaptation (QPA) in their study on quantum circuit-based compression for parameter-efficient learning. QPA leverages quantum neural networks (QNNs) to generate classical model weights during training, decoupling inference from quantum hardware. This approach significantly reduces parameters in fine-tuning methods like Low-Rank Adaptation (LoRA) for models such as GPT-2 and Gemma-2. Specifically, QPA reduces parameters to 52.06% for GPT-2 with a 0.75% performance gain, and to 16.84% for Gemma-2 with a 0.07% improvement. These findings underscore the potential of quantum-enhanced parameter reduction, offering scalable solutions for fine-tuning large language models while maintaining classical hardware feasibility.
https://arxiv.org/abs/2410.09846v1
A new metric to help find quantum advantage for ground state problems
IBM researchers, led by Antonio Mezzacapo and Javier Robledo-Moreno, have introduced a new metric, the "V-score," to benchmark quantum computing's ability to solve ground state problems, as published in *Science*. This metric evaluates the accuracy of quantum algorithms in estimating the ground state energy of systems, a challenging task with applications in physics and chemistry. The study involved 29 institutions and tested the V-score on the largest set of local Hamiltonian problems to date. The V-score helps identify problems where quantum computing could surpass classical methods, marking potential quantum advantage areas. As quantum algorithms evolve, the V-score will be crucial in assessing their effectiveness and identifying opportunities for quantum breakthroughs.