In this blog post, we’ll explore the principles and characteristics of memristors—electronic devices that mimic human synapses—and their significance in artificial intelligence technology.
When you first ride the subway, the clattering sound of the wheels seems unusually loud. However, as the train passes one station after another, there comes a point when we barely notice that sound anymore. This is because our senses have become desensitized as the same stimulus is repeated. In contrast, a machine’s senses do not easily become desensitized. Let’s take a beverage vending machine as an example. The machine accurately detects which button has been pressed and then dispenses the corresponding drink. Even if this process is repeated all day long, the machine performs the same action without hesitation the next day. For most machines, performance remains consistent even during such repetitive tasks. So, what if machines possessed the characteristic of having their senses dulled, just like humans?
Sensory dullness is closely related to synapses. Synapses are the connections between neurons—the nerve cells that make up the human brain—and sensory information is transmitted in the form of electrochemical signals as it passes through neurons and synapses. When the same signal is transmitted repeatedly, the efficiency of neural transmission at the synapse temporarily decreases, resulting in a phenomenon known as sensory dullness. So, is it possible to mimic these characteristics of synapses in computers?
The core of conventional computers consists of countless transistors. In fact, research was previously conducted to mimic human synapses using transistors. However, it is known that the human brain contains over 100 trillion synapses, and implementing all of them with transistors faces significant limitations in terms of integration density and power consumption. As transistor-based neural networks grow in scale, their space requirements and power consumption increase dramatically. For this reason, researchers are turning their attention to a new electronic device called the memristor, which can implement characteristics more similar to those of human synapses.
The term “memristor” is a portmanteau of “memory” and “resistor”; it refers to an ultra-small electronic device that retains the influence of past current flow. While a resistor typically determines the magnitude of current flow, a memristor’s resistance value changes depending on the total amount of current that has flowed through it. It also retains its last resistance state even after the current supply is cut off. These characteristics are considered a major advantage because they allow information to be stored without the need for a separate memory device. While conventional transistors can also store information, they require additional circuitry, which increases the size of the device and raises power consumption. In contrast, memristors have a simple structure, consume little power, and are well-suited for high integration; as a result, they are being researched as a key technology in the fields of next-generation memory and neuromorphic computing.
Although the concept of the memristor was first proposed by Professor Leon O. Chua in 1971, it did not receive much attention at the time because transistor technology was advancing rapidly. However, after researchers at Hewlett-Packard successfully implemented a real memristor in 2008, related research gained significant momentum, and it is now regarded as one of the core technologies in the fields of next-generation semiconductors and artificial intelligence hardware.
A memristor is fundamentally structured with titanium oxide (TiO₂) sandwiched between platinum electrodes. Titanium dioxide is divided into two regions with different properties depending on the distribution of oxygen. One region is titanium dioxide (TiO₂), which is fully saturated with oxygen and where oxygen hardly moves; the other is titanium dioxide (TiO₂-x), where some oxygen is missing, allowing oxygen to move relatively freely. Since current flows easily through the region where oxygen is partially depleted even at low voltages, the overall electrical resistance decreases as the proportion of this region increases. Platinum electrodes are placed above and below the titanium dioxide layer to apply voltage.
Now, let’s consider the case where a negative voltage is applied to the upper electrode. Since like charges repel each other, the oxygen ions near the interface move toward the opposite region. As a result, part of the region that was previously saturated with oxygen becomes a region where electricity flows easily as oxygen escapes, and the region where current flows easily gradually expands. Conversely, if a negative voltage is applied to the lower electrode, the oxygen ions move back in their original direction, causing the region with high electrical conductivity to shrink, while the region with sufficient oxygen expands again. These changes continue as long as the voltage is applied, demonstrating the first characteristic of a memristor: its resistance value changes depending on the amount of current flowing.
So why does the resistance value remain unchanged even after the power is turned off? Transistors also change their state depending on voltage, but they generally cannot maintain that state once the power is cut off. In contrast, since the atomic arrangement within the titanium oxide itself changes in a memristor, the final state is retained even after the power is cut off.
In regions where oxygen is partially deficient, two types of titanium ions (Ti³⁺, Ti⁴⁺) exist. When oxygen ions move, electrons shift between titanium ions to maintain charge balance, and in the process, some Ti³⁺ ions are converted into Ti⁴⁺ ions. Since Ti⁴⁺ ions, which carry a greater positive charge, attract negatively charged oxygen ions more strongly, this charge transfer occurs naturally. Ultimately, even after the voltage is removed, the new arrangement of atoms remains unchanged, so the resistance also maintains its final state.
To put this in simple terms, it’s similar to a situation where one person is sitting in a subway seat and the seat next to them is empty. Every time the train moves, the person shifts one seat over to the empty seat, but the number of empty seats remains at one. Just as the person’s position changes but the overall structure of the seats remains the same, in a memristor, only the arrangement of oxygen ions and electrons changes while the overall structure is preserved, allowing it to “remember” its previous state.
Scientists are noting that this characteristic is very similar to that of human synapses. In synapses, the efficiency of neural transmission changes as the same signal is repeated; similarly, in memristors, the resistance value changes when current flows repeatedly. Because of these similarities, memristors are being researched as core components of neuromorphic computing, which mimics the human brain. In fact, researchers are developing various types of artificial intelligence hardware by combining memristors with CMOS (Complementary Metal Oxide Semiconductor) circuits to mimic the behavior of neurons and synapses, and related research is currently underway.
In a few decades, even beverage vending machines might be equipped with artificial intelligence that resembles the human brain. Imagine standing in front of a vending machine and pressing a button, only for the machine to recognize your fingerprint or other biometric information and say, “I know what drink you had last summer.” You might just reply, “Then give me the same one as last time.”
