Chinese Researchers Fine-Tune a Billion-Parameter AI Model using a Quantum Computer
According to the state-affiliated Global Times, Chinese researchers have refined a billion-parameter artificial intelligence (AI) model using a quantum computer, which they say is a first for the world and a step towards integrating quantum computing with complex AI tasks.
Origin Wukong, a superconducting quantum computer powered by a 72-qubit chip manufactured in the country, was used for the experiment. The fine-tuning work demonstrated that quantum hardware might enhance model training efficiency even when the number of model parameters was dramatically reduced, providing a potential remedy for the growing demands on computing capacity, according to Global Times.
Anhui Engineering Research Centre for Quantum Computing
The Anhui Quantum Computing Engineering Research Centre, which runs Origin Wukong, announced the development Monday. According to the center, a single batch of input data can initiate hundreds of quantum processes at once, allowing for large-scale parallel processing.
According to the publication, the scientists used Origin Wukong to refine a billion-parameter model in one trial. The resulting model demonstrated a 15 percent reduction in training loss, a crucial indicator of a model’s learning performance, using a dataset of psychological counselling dialogues. The accuracy of the model increased from 68 percent to 82 percent in a different challenge that tested mathematical reasoning.
Even after cutting the model size by 76%, researchers observed an 8.4% increase in training effectiveness, according to the Global Times, a publication connected to China’s state-run People’s Daily. According to scientists, this could result in more effective models that require less energy and memory to operate.
A typical stage in the development of AI is fine-tuning, which is used to modify large language models (LLMs) that have already been trained for certain applications, such as customer service or legal advice. Adjusting millions or billions of parameters internal numerical values the model uses to make predictions usually takes a lot of processing power. The workload that is typically handled by classical computers may be reduced if this is done on quantum technology.
A Significant Advancement
Chen Zhaojun, a deputy researcher at the Institute of Artificial Intelligence under the Hefei Comprehensive National Science Centre, told Global Times that this was a significant advancement in the realm of quantum computing. He informed the publication that this was the first instance of an actual quantum computer being used to assist with fine-tuning a huge model, demonstrating that the technology is capable of handling such jobs.
According to Dou Menghan, deputy director of the Anhui Quantum Computing Engineering Research Centre and vice president of Origin Quantum Computing Technology Co., the endeavour was successful by combining quantum tools with classical systems. “It’s similar to adding a ‘quantum engine,’ to a classical model,’ enabling the two to function in tandem,” Dou told Global Times.
The fine-tuning task was performed by the researchers using Origin Wukong over a sequence of quantum computing jobs. In theory, quantum systems are well suited for concurrent model training, which each job was able to sustain. Qubits, which are units of quantum information, are used by quantum computers to carry out calculations that are challenging or impossible for classical systems.
According to Global Times, Origin Wukong’s 72-qubit chip is one of the most potent devices ever created in China and is regarded as one of the most sophisticated systems globally. Since its launch on January 6, 2024, Wukong has finished over 350,000 computations for applications in biomedicine, finance, and fluid dynamics. According to information given to Global Times, users in 139 nations and regions have also requested remote access.
The name Wukong comes from Chinese mythology’s Monkey King, Sun Wukong. The figure was selected to represent the machine’s versatility and breadth across industries and problem types because of his well-known capacity to change into 72 various shapes.
According to industry experts quoted by Global Times, the trial might provide a solution to the “computing power anxiety” that permeates the AI space. The demands on hardware increase as LLMs get bigger, which raises expenses and energy consumption. Even in their early phases, quantum processors have the potential to eventually assist counterbalance such needs.
Global rivalry is still fierce despite China’s significant spending in quantum research and hardware. Programs to combine quantum and AI have also been started in the US, Europe, and Canada. However, Origin Wukong is one of the first quantum platforms to manage a full-scale AI fine-tuning workload with measurable outcomes, said the researchers behind this most recent testing.
It’s crucial to remember that the project is still merely a demonstration and not a complete commercial use. The report makes no mention of the researchers publishing or planning to publish a study on the experiment, which would enable a more in-depth examination of the experiment.
