Summary: Researchers have created a brain-inspired semiconductor device that reproduces a neuron’s ability to change its sensitivity based on experience—known as intrinsic plasticity. The “Frequency Switching Neuristor” combines two complementary memristor types to control spike frequency, allowing the artificial neuron to learn from past activity and autonomously adapt its response. This hardware-level plasticity cuts energy usage and increases robustness, pointing to more efficient, fault-tolerant AI hardware for edge computing, autonomous vehicles, and other real-world applications.
In simulation, systems using this neuristor achieved equal task performance while reducing energy consumption by 27.7% versus conventional neural-network implementations. The device also demonstrated structural resilience: when some artificial neurons were damaged, the network reorganized through intrinsic plasticity and recovered its performance. These results suggest a path toward neuromorphic hardware that is both energy-efficient and self-stabilizing.
Key Facts
- Intrinsic plasticity: The device reproduces a neuron’s capacity to adjust excitability based on recent activity.
- Energy efficiency: Simulated networks reached the same accuracy with approximately 27.7% lower energy consumption.
- Self-recovery and resilience: Built-in plasticity enabled networks to reconfigure and restore function after simulated neuron damage.
Source: KAIST
Background: Beyond synaptic weight changes, biological neurons exhibit intrinsic plasticity—their own adjustable excitability that makes them more or less responsive depending on prior experience. This mechanism helps the brain filter repeated, irrelevant inputs and to become sensitized through training. Most current AI chips emulate synaptic learning but do not natively capture this form of neuronal adaptability.
A team at KAIST led by Professor Kyung Min Kim (Department of Materials Science and Engineering) developed the Frequency Switching Neuristor to embed intrinsic plasticity directly into a single semiconductor element. The device integrates a volatile Mott memristor—one that momentarily switches state and then returns to its baseline—with a non-volatile valence change memory (VCM) memristor that retains state over longer periods. Together, these elements let the neuristor program and adjust spiking frequency in response to prior inputs, effectively combining memory and processing inside one unit.
Functionally, spike generation and resistance changes in the paired memristors interact: recent activity changes the device’s transfer characteristics so that future spikes become more or less frequent. This mirrors how biological neurons can become habituated to recurring stimuli or progressively sensitized during repeated training.
The researchers validated the approach using device-based simulations of sparse neural networks. Because each artificial neuron carries its own short-term memory and adjustable excitability, the network exploited these local dynamics to perform tasks while consuming substantially less energy. The simulations showed a 27.7% reduction in energy for comparable performance versus conventional approaches that lack intrinsic plasticity.
Beyond efficiency gains, the device-enabled networks exhibited structural plasticity: when a fraction of nodes were randomly disabled, intrinsic plasticity allowed the remaining network to reconfigure and recover full performance. This built-in ability to adapt to partial hardware failure is particularly valuable for deployed systems—such as edge devices and autonomous vehicles—that demand long-term stability and graceful degradation under real-world stress.
Professor Kyung Min Kim commented that implementing intrinsic plasticity in a single semiconductor element advances both energy efficiency and reliability for AI hardware. By letting devices remember and adapt their own internal state, the technology can help create neuromorphic systems that sustain operation over long periods and recover from damage without centralized intervention.
The study lists Dr. Woojoon Park (now at Forschungszentrum Jülich, Germany) and Dr. Hanchan Song (now at ETRI) as co-first authors. The results were published online on August 18 in the journal Advanced Materials (IF 26.8).
Funding: This work was supported by the National Research Foundation of Korea and Samsung Electronics.
About this neurotech research news
Author: JEEHYUN LEE
Source: KAIST
Contact: JEEHYUN LEE – KAIST
Image: Image credit: Neuroscience News
Original Research: Open access. “Frequency Switching Neuristor for Realizing Intrinsic Plasticity and Enabling Robust Neuromorphic Computing” by Kyung Min Kim et al., Advanced Materials.
Abstract
Frequency Switching Neuristor for Realizing Intrinsic Plasticity and Enabling Robust Neuromorphic Computing
Biological information processing owes much of its efficiency and flexibility to spatiotemporal spiking activity combined with intrinsic plasticity—the autonomous adjustment of neuronal excitability based on prior stimulation. Mott memristors, which display threshold switching, are promising building blocks for artificial neurons (neuristors) that produce spiking behavior. Until now, systematic implementation and evaluation of intrinsic plasticity within neuromorphic devices have been limited.
The Frequency Switching (FS) neuristor introduced here emulates intrinsic plasticity by pairing a volatile Mott memristor with a non-volatile VCM memristor. This combination creates programmable, multi-level frequency–voltage (f–V) transfer characteristics analogous to neuronal intrinsic plasticity functions. Device-level simulations of sparse neural networks indicate that intrinsic plasticity serves both as memory and local processor, improving overall network efficiency and lowering energy demands. Furthermore, the mechanism provides structural plasticity, enabling full performance recovery after random neuron loss and suggesting a route toward more adaptive, energy-efficient, and fault-tolerant neuromorphic computing systems.