Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have developed a self-learning memristor for neuromorphic computing that can correct its own errors, a development detailed in the journal Nature Electronics.
The theoretical basis for this technology dates to 1971, when American electrical engineer and computer scientist Leon Chua proposed the existence of a fourth fundamental electrical element. Alongside the resistor, capacitor, and inductor, Chua reasoned there must also be a “memristor,” a term created by combining “memory” and “resistor.” He described this component as possessing non-volatile memory, meaning it could retain stored information even without a power source. This concept laid the groundwork for future advancements in memory components that could more closely replicate biological brain functions.
While the memristor was a theoretical construct for decades, its experimental discovery by researchers in 2008 generated significant interest within the scientific community. Following this discovery, memristors became a leading candidate for use as artificial synapses in neuromorphic, or brain-like, computing systems. Their key attribute is the ability to perform both data storage and computation simultaneously within a single component. This dual functionality mirrors the integrated way synapses operate in the human brain, where processing and memory are not physically separated, offering a path toward creating more efficient and powerful artificial neural networks.
In January 2024, KAIST president Kwang Hyung Lee announced a successful development in this field. The institute’s researchers created a memristor that can learn from its mistakes and correct errors, enabling it to address problems previously considered difficult for neuromorphic systems. As a specific example of its capability, the research team stated the new chip can execute tasks like separating a moving image from a static background during video processing. The system is designed to improve its accuracy at this task as it continues to perform it over time.
This advancement facilitates the local performance of artificial intelligence tasks, reducing the reliance on remote cloud-computing servers. This shift in processing location directly enhances user privacy by keeping data on the device and improves overall energy efficiency. Hakcheon Jeong and Seungjae Han, researchers from KAIST, provided an analogy for the system’s operation in a press statement. “This system is like a smart workspace where everything is within arm’s reach instead of having to go back and forth between desks and file cabinets,” they said. “This is similar to the way our brain processes information, where everything is processed efficiently at once at one spot.”
In a related development during the same year, KAIST also introduced an AI superconductor chip. This chip was designed to operate at ultra-high speeds while consuming minimal power, another attribute that emulates the efficiency of the human brain. For context, the human brain is capable of performing approximately a billion-billion, or 10^18, mathematical operations per second while using only about 20 watts of power. Creating neuromorphic systems that can achieve this level of hyper-efficiency is a central goal of the field.
The continued improvement of memristors is considered an incremental step toward the creation of a true brain-on-a-chip, a technology that could accelerate AI capabilities. This progress has led to discussions about the technological singularity, a hypothetical point when artificial intelligence surpasses human intelligence. However, the complexity of the term “intelligence” is noted by scientists, who point out that the ability to perform calculations like a human brain does not equate to possessing all of the brain’s functions.
Some scientific perspectives suggest that such machines could constitute “alien minds,” with neural constructions that are different from human cognition but represent their own form of intelligence. At present, the human brain remains the benchmark for hyper-efficient computing, though advancements in components like memristors may enable AI to achieve similar performance levels.
