The Korea Advanced Institute of Science and Technology (KAIST) announced in January 2024 the development of a self-learning memristor, a component designed to replicate the function of synapses in the human brain.
According to KAIST President Kwang Hyung Lee, the new device can correct its own errors and improve its performance over time, addressing previous challenges in neuromorphic systems. The research, published in the journal Nature Electronics, outlines the memristor’s capabilities. Researchers report that the chip can, for example, learn to separate a moving image from its background during video processing and progressively enhance its ability to perform this task.
This advancement could allow for complex AI tasks to be executed locally on devices rather than relying on remote cloud servers, which would increase both privacy and energy efficiency. “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,” said KAIST researchers Hakcheon Jeong and Seungjae Han in a press statement.
“This is similar to the way our brain processes information, where everything is processed efficiently at once at one spot.” The memristor, a term derived from “memory” and “resistor,” is considered a foundational element for neuromorphic, or brain-like, computing. The concept was first theorized in 1971 by American electrical engineer and computer scientist Leon Chua.
He proposed that a fourth fundamental electrical component must exist alongside the resistor, capacitor, and inductor. Chua envisioned the memristor as a non-volatile memory component capable of storing information even when powered off. Although the theory existed for decades, researchers did not experimentally discover memristors until 2008.
This breakthrough sparked global scientific efforts to improve their capabilities. A memristor’s ability to perform both data storage and computation simultaneously makes it an effective stand-in for an artificial synapse in an AI neural network, mimicking how the human brain functions. A primary goal of this research field is to build computers that can operate with the efficiency and power of the human brain.
The brain can perform an estimated one billion-billion (10^18) mathematical operations per second using just 20 watts of power. Reaching this level of hyper-efficiency is a key requirement for developing a practical neuromorphic AI brain. In a related development this year, KAIST also created the first AI superconductor chip.
This chip is designed for ultra-high-speed operation with minimal power consumption, further emulating the efficiency of the brain. These technological improvements are viewed as incremental steps toward creating a “brain-on-a-chip.” Such technology could significantly advance AI and potentially accelerate progress toward the singularity, a theoretical future point where artificial intelligence surpasses human intelligence.
However, the article notes that “intelligence” is a complex subject. An AI’s capacity to perform certain calculations similar to the human brain does not mean it can replicate all of the brain’s diverse functions. Some scientists speculate that such machines might evolve into “alien minds,” possessing neural constructions that are intelligent in a way that is fundamentally different from human cognition.
For the present, the human brain remains the standard for hyper-efficient computing. Through continued advancements with components like memristors, AI may eventually challenge that position.
