Dr Li Can, an assistant professor of our department, has won this year’s Croucher Tak Wah Mak Innovation Award with a fund of HK$5 million with help of his research in emerging memory devices.
Dr Li said that training AI models based on existing computer systems would require a huge amount of time and energy. Taking ChatGPT as an example, he mentioned that the model takes 36 years to train using eight state-of-the-art graphic processing units for a GPT-3 model, and even longer for the latest GPT-4 model. His team found that the memristor, which mimics the behavior of biological synapses and neurons in human brains like tolerating defects, learning from experience, and reasoning based on vague information, can unlock the potential of training a human-like AI in a more power-efficient way. The computing speed of brain-inspired memristive hardware could be 100 to 1,000 times faster than the current AI model. Li’s team has already produced a chip model to verify the feasibility of computing paradigms. 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 and his ultimate goal is to create a chip just like “a newborn with unique DNA” that can think and even sense.
Dr Li hopes to recruit more young people to his team and achieve acceleration of the computing paradigms in the next three to five years.
