Ayra Wang

Brain-inspired computing can generate more powerful artificial intelligence that conventional computers cannot achieve, a University of Hong Kong scientist said.

Research in emerging memory devices, such as a memristor, has helped the assistant professor of the department of electrical and electronic engineering, Li Can, to win this year's Croucher Tak Wah Mak Innovation Award with a fund of HK$5 million.

Li said training AI models based on existing computer systems will require a huge amount of time and energy.

"It will need a nuclear power plant to power the computing system of the AI model after 10 years if we keep using conventional computers to train AI," Li said.

Taking ChatGPT as an example, he said it took researchers 36 years on eight processing units to train a GPT-3 model.

It would be even more time-consuming and require more electric power to train a more advanced AI model, he said.

Li said conventional computers will also be obstacles to the emergence of creativity in AI as they can only learn from limited databases and follow precise instructions.

"Conventional computers are not an optimal solution to creating AI," Li said.

His team found that the memristor, which mimics the behavior of biological synapses and neurons in human brains, can unlock the potential of training a human-like AI in a more power-efficient way.

"Human brains can tolerate defects, learn from experience, and reason based on vague information," Li said, adding that those features of human brains could be key functions of an advanced AI that can learn lifelong and have creativity.

Li said brain-inspired memristive hardware could compute directly within memory, which can save time and energy consumed during the process of transporting data from storage to the computing unit in conventional computers.

"The computing speed of brain-inspired memristive hardware could be 100 to 1,000 times faster than the current AI model," Li said.

But his team aims to create new brain-inspired computing paradigms that could speed up computing even more, and create an integrated memristor chip that would consume less energy.

"The future trend of AI development will be internet-free, as AI models will be put in and can run in terminals such as phones, or even a small chip that can be implanted into human bodies that can help detect diseases," Li said.

His ultimate goal is to create a chip just like "a newborn with unique DNA".

"Every chip would be like a new life that can think and even sense," Li said, adding that various AI models designed for chips could shape their different functions.

Li's team has already produced a chip model to verify the feasibility of computing paradigms, and he hopes to recruit more young people to his team and achieve acceleration of the computing paradigms in the next three to five years.

Li Can

Memory device aims to boost AI