Hong Kong researchers seek key to advanced AI chip within original model – the human brain

“Our brain can tolerate defects, meaning it could work despite the presence of dead cells. This is fundamentally different from computers that cannot work with a malfunctioning transistor,” Li said, referring to the primary building block of microchips, including the central processing unit (CPU).

“We can learn from experience, unlike computers and powerful AI. [A chatbot like] ChatGPT might tell you that it does not know about the latest news because of its knowledge cut-off at a certain time. Humans are also able to reason based on vague information, while computers require clear instructions.”

But there is an area where computers have an edge. “Computers are more suitable for performing scientific calculations and repetitive operations as well as deriving values from certain patterns than the human brain,” Li, who has been working in the field for more than a decade, said.

In their quest to help computers better emulate the human brain, Li and his team have been looking at an emerging memory device. Called a memristor, the advanced microelectronic platform can copy the behaviour of biological synapses and neurons, the pathways of the human brain.

“Each chip is like a newborn baby with DNA that determines its traits. It would be a training model for a chip. But how a baby develops as it grows remains to be seen, and the same goes for a chip.”

“Brain-inspired devices we are working on would result in a technology different from conventional computers. It will also be a promising addition to speeding up graphics processing units [GPUs],” he said, referring to a computer chip used to train AI systems.

According to US semiconductor giant Nvidia, training a GPT-3 large language model with 175 billion parameters would take 36 years on eight GPUs, or seven months with 512 GPUs.

GPT-3 was the third-generation large language model released by ChatGPT creators OpenAI in 2020. The latest model in use is the fourth version – GPT-4.

A 2021 preprint paper by a team of researchers from Google and the University of California, Berkeley, posted ahead of peer review, estimated that training the model took 1,287 megawatt hours, equivalent to 120 years of electricity consumption for an average American household.

“With AI models expanding to beyond 1 trillion parameters, transferring data would pose a bottleneck. The memory system will become increasingly important,” Li said, adding that more energy-efficient hardware would be required to support increasing computational needs.

This is where the memristor comes into play. In mimicking the action of the human brain, it is able to compute directly within memory, removing the need for data transfers between memory and processing units, just as the brain processes information where it is stored, Li said.

Chinese scientists unlock potentials of a new semiconductor building block

He said future memristor chips could potentially offload AI tasks from energy-intensive data centres to devices such as smartphones and watches with great energy efficiency.

These could also find application as subdermal or under-the-skin health-monitoring implants and speed up virus genome sequencing, according to Li.

“To make a wearable sensor to monitor diseases, for example, the device needs to be highly energy efficient without compromising its functionality – hopefully it could run years on a single charge,” he said.

He said such chips would also have the potential to reduce the time needed to sequence a virus, which currently takes a few days to a week. The coronavirus that causes Covid-19, for example, carries its genetic information in a single strand of RNA consisting of 30,000 letters of genetic code.

Earlier this month, Li received HK$5 million (US$640,000) in funding as a recipient of this year’s Croucher Tak Wah Mak Innovation Awards. The grant is sponsored by the Croucher Foundation, a Hong Kong-based private organisation that supports local scientists.

Li said half of the funds would go towards hiring new team members, and the rest used to support experiment expenses and explore cutting-edge research.

“Hong Kong and HKU are in the position to attract top talent. Funding is also critical,” he said.