⚡ electrical engineering

A Non-Monotonic Reconfigurable Neural Unit with Dynamic Integration Control for Advanced Neuromorphic Computing

This paper introduces a reconfigurable neural unit (RNU) that integrates a MoTe2/h-BN synaptic transistor with an Ag/h-BN/graphene volatile memristor to achieve a dynamically tunable U-shaped non-monotonic activation function and precise spatiotemporal inhibitory control, thereby enhancing the performance of neuromorphic systems in image recognition and EEG decoding tasks.

Original authors: Jing Liu, Xitong An, Yan Wang, Chao Dou, Haoyue Lu, Yueying Li, Xuan Deng, Dong Sun

Published 2026-06-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jing Liu, Xitong An, Yan Wang, Chao Dou, Haoyue Lu, Yueying Li, Xuan Deng, Dong Sun

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a computer that thinks more like a human brain. The problem is, most current "brain-like" computer chips are a bit too rigid. They act like simple light switches: if you push the button (input), the light turns on (output). If you push harder, the light gets brighter. They can't really "think" in the nuanced, dynamic ways biological neurons do, such as ignoring certain signals or changing their behavior based on the situation.

This paper introduces a new, smarter building block for these computers called a Reconfigurable Neural Unit (RNU). Think of it as a "smart switch" that can change its own rules on the fly.

Here is how it works, broken down into simple concepts:

1. The Two-Part Team

The researchers built this unit by stacking two different materials on top of each other, creating a tiny sandwich:

The "Leaky Bucket" (The Memristor): This part acts like a bucket with a hole in the bottom. When you pour water (electrical signals) in, the level rises. If you pour fast enough, it overflows (fires a signal). But if you stop pouring, the water leaks out. This mimics how real neurons wait for enough signals before "firing."
The "Smart Gatekeeper" (The Transistor): This is the special part. It sits in front of the bucket and controls how easily the water can get in. Unlike a normal gate that just opens wider when you push, this gate is ambidextrous (it works both ways). It creates a strange "U-shape" behavior:
- If you push the gate slightly left, it opens.
- If you push it slightly right, it opens.
- But if you push it right in the middle, it slams shut completely, blocking all water.

This "middle silence" is a big deal. In the brain, this is like lateral inhibition—ignoring a specific type of noise while listening to everything else. It allows the computer to be selective about what it pays attention to.

2. The "Magic Light" Tuning Knob

One of the coolest features is that this "silent middle" isn't stuck in one place. The researchers found a way to move it using ultraviolet (UV) light.

Imagine the "silent zone" is a dark spot on a stage where no actors can perform.
By shining a specific color of UV light on the device, they can physically shift that dark spot to the left or right.
This means they can reprogram the neuron's behavior instantly. If the computer needs to focus on a different type of signal, they just shine a light, move the "silent zone," and the neuron adapts to the new task.

3. The "Stop Sign" Button

The device also has a special input called a "gate" that acts like a dedicated Stop Sign.

Normally, the neuron waits for signals to pile up to fire.
But if this special "Stop Sign" button is pressed (using an inhibitory signal), it cuts the process short. It prevents the neuron from firing even if it was about to.
This gives the system precise control over when to fire and when to stay quiet, which is crucial for processing complex, time-based information.

4. Did It Work? (The Test Drive)

The researchers tested this new "smart switch" in two scenarios:

The Photo Test (Fashion-MNIST): They asked the system to recognize pictures of clothes (like shirts vs. coats). The new system performed almost as well as standard, non-brain-like computers, proving it can handle basic image recognition.
The Brainwave Test (EEG): They asked the system to decode human brainwaves to guess what a person was imagining (like moving their left hand vs. right hand). This is much harder because brainwaves change rapidly over time.
- The Result: The system with the "Stop Sign" and "Smart Gate" (the RNU) was 3.5% more accurate than standard brain-like computers. This proves that the ability to dynamically suppress signals and change rules on the fly makes a real difference when dealing with messy, real-time data.

The Bottom Line

This paper presents a tiny, energy-efficient device that behaves much more like a real biological neuron than current technology. It can ignore specific signals, change its own sensitivity using light, and act as a "stop sign" for information. By using this building block, future computers could be better at handling complex, time-sensitive tasks like reading brainwaves or adapting to new environments, all while using very little power.

Technical Summary: A Non-Monotonic Reconfigurable Neural Unit with Dynamic Integration Control

Problem Statement
Current neuromorphic hardware implementations face fundamental limitations in emulating the rich, adaptive dynamics of biological neural systems. Most existing artificial neurons rely on fixed, monotonic activation functions and lack the ability to dynamically control temporal integration processes. Specifically, they fail to reproduce biological mechanisms such as lateral inhibition (which creates feature-selective "silent regions" to suppress specific input ranges) and context-dependent modulation of firing thresholds. Furthermore, existing device architectures often integrate leaky integrate-and-fire (LIF) dynamics and firing modulation into a single active channel, creating a trade-off where tuning one function compromises the other.

Methodology
To address these constraints, the authors propose a Reconfigurable Neural Unit (RNU) based on a van der Waals heterostructure that decouples neuronal functions into two specialized components:

Volatile Memristor (Ag/h-BN/graphene): This component serves as the core for LIF dynamics, providing volatile switching necessary for neuronal firing and self-resetting behavior.
Ambipolar Synaptic Transistor (MoTe2/h-BN): This component is connected in series with the memristor to enable sophisticated firing modulation.

The device architecture leverages the unique properties of the MoTe2 channel:

Non-Monotonic Response: By applying gate biases, the conductance of the MoTe2 channel is modulated. Due to its ambipolar nature, the channel exhibits a minimum conductance point. When the gate bias is near this minimum, the transistor enters a high-resistance state, suppressing voltage across the memristor and creating a "silent region" where firing is inhibited. As the gate bias moves away from this point in either direction, conductance increases, allowing firing to resume. This results in a U-shaped, non-monotonic firing response.
Dynamic Reconfiguration: The position of this silent region on the gate-bias axis is dynamically reconfigurable via ultraviolet (UV)-light-assisted doping. UV illumination (365 nm or 254 nm) combined with specific gate biases induces n-type or p-type doping in the MoTe2 channel, permanently shifting the conductance minimum and thus reconfiguring the RNU's operational window.
Inhibitory Control: The gate terminal of the transistor functions as a dedicated input for inhibitory spikes, allowing for precise spatiotemporal control over the integration window and firing dynamics.

Key Contributions

Device Architecture: The successful fabrication and characterization of a dual-component RNU that physically separates synaptic tuning (transistor) from firing dynamics (memristor), overcoming the limitations of single-channel designs.
Non-Monotonic Excitability: Demonstration of a U-shaped firing response with a central silent region, mimicking biological lateral inhibition and enabling feature-selective suppression.
Adaptive Tunability: Proof of concept for non-volatile, UV-assisted reconfiguration of the neuronal transfer function, allowing the "silent region" to be shifted to adapt to different computing tasks.
Inhibitory Pathway: Implementation of a feedforward inhibition mechanism where the gate terminal directly modulates temporal integration, truncating integration windows to suppress firing.

Results

Device Characterization:
- The MoTe2/h-BN transistor demonstrated balanced ambipolar behavior with an on/off ratio of $10^5$ and a minimum drain current of 12.929 pA.
- UV doping successfully shifted the conductance minimum from 2 V (intrinsic) to -10 V (n-doped) and +22 V (p-doped).
- The Ag/h-BN/graphene memristor exhibited reproducible volatile switching with a threshold voltage of ~~0.18 V and low leakage current (~~1 pA).
- The integrated RNU achieved an average power consumption of 0.04 nW per spiking event.
Neural Dynamics:
- The RNU exhibited amplitude-dependent LIF behavior and a gate-tunable non-monotonic firing profile.
- Inhibitory gate signals successfully truncated temporal integration windows, suppressing firing for subsequent inputs.
System-Level Performance:
- Static Image Recognition (Fashion-MNIST): Spiking Neural Networks (SNNs) equipped with RNU dynamics achieved 90.42% accuracy, comparable to conventional Artificial Neural Networks (91.62%) and demonstrating the effective transfer of device-calibrated dynamics to static tasks.
- Dynamic EEG Decoding: In a four-class motor imagery EEG decoding task, a Gate-Controlled SNN (GCSNN) incorporating the RNU's inhibitory pathway achieved 64.6% accuracy. This represents a 3.5% absolute improvement over a standard SNN baseline (61.1%) and a significant gain over a conventional CSP+LDA baseline (45.1%). The GCSNN showed superior performance in distinguishing temporal patterns and reducing inter-class confusion.

Significance
The paper posits that the RNU provides a compact, versatile, and highly adaptive foundation for advanced neuromorphic systems. By emulating biological non-monotonic responsivity and enabling dynamic reconfiguration of neuronal transfer functions, the RNU addresses the rigidity of current hardware. The authors claim that this architecture supports more biologically plausible and computationally powerful paradigms, such as adaptive sensory filtering and selective attention. The results suggest that RNU-based systems are particularly well-suited for tasks requiring dynamic temporal regulation and processing of non-stationary bio-signals, marking a significant step toward the hardware realization of truly adaptive artificial intelligence.