Brain-Inspired Memory Device Aims to Boost AI
Brain-inspired computing holds the key to unlocking more powerful artificial intelligence (AI) capabilities that conventional computers cannot achieve, according to a scientist from the University of Hong Kong. Dr. Li Can, an assistant professor in the department of electrical and electronic engineering, believes that research in emerging memory devices, such as the memristor, can revolutionize AI training and unleash its full potential.
Dr. Li’s groundbreaking work in this field has earned him the prestigious Croucher Tak Wah Mak Innovation Award, along with a HK$5 million fund. He emphasized that training AI models using existing computer systems demands an exorbitant amount of time and energy, to the point where it could require a nuclear power plant to sustain the computing system over the next decade.
To illustrate the challenges posed by conventional computers, Dr. Li referred to the example of ChatGPT, highlighting that the training of a GPT-3 model using eight processing units took a staggering 36 years. He further emphasized that training more advanced AI models would be even more time-consuming and energy-intensive.
One of the limitations of conventional computers lies in their inability to foster creativity in AI, as they can only learn from limited databases and follow specific instructions. Recognizing this, Dr. Li and his team turned to the memristor, a memory device that mimics the behavior of synapses and neurons in the human brain. They found that leveraging memristive hardware could facilitate the training of human-like AI in a more power-efficient manner, enabling the integration of defects tolerance, experiential learning, and reasoning based on vague information, all of which are key features of the human brain.
Moreover, brain-inspired memristive hardware has the potential to compute directly within memory, eliminating the need to transport data from storage to the computing unit in conventional computers. According to Dr. Li, this advancement could significantly speed up computing, with brain-inspired memristive hardware boasting computing speeds 100 to 1,000 times faster than current AI models.
Dr. Li’s ultimate vision is to develop an integrated memristor chip that consumes minimal energy and can be deployed in a wide range of applications. He envisions a future where AI models can operate independently without relying on internet connectivity, making it possible to embed them in various devices such as smartphones or even small chips implanted in human bodies for disease detection.
Comparing each chip to a newborn with unique DNA, Dr. Li sees the potential for different AI models to shape their functions, essentially creating a chip that can think and sense. To achieve this ambitious goal, he plans to expand his team and accelerate the development of computing paradigms over the next three to five years.
In summary, brain-inspired computing utilizing memristive hardware holds immense potential for advancing AI capabilities far beyond the limitations of conventional computers. Its ability to emulate the human brain’s creative thinking, tolerance for imperfections, and reasoning based on incomplete information brings us one step closer to developing a new generation of powerful and efficient AI systems. With further research and innovation, Dr. Li Can and his team hope to usher in a new era of intelligent machines that can revolutionize various aspects of our lives.
