NbO2-based memristive neurons for burst-based perceptron

Neuromorphic computing using spike-based learning has broad prospects in reducing computing power. Memristive neurons composed with two locally active memristors have been used to mimic the dynamical behaviors of biological neurons. In this work, the dynamic operating conditions of NbO2-based memristive neurons and their transformation boundaries between the spiking and the bursting are comprehensively investigated. Furthermore, the underlying mechanism of bursting is analyzed and the controllability of the number of spikes during each burst period is demonstrated. Finally, pattern classification and information transmitting in a perceptron neural network by using the number of spikes per bursting period to encode information is proposed. The results show a promising approach for the practical implementation of neuristor in spiking neural networks.

💡 Research Summary

The paper presents a comprehensive study of neuromorphic hardware based on NbO₂‑based memristive neurons that can operate in both spiking and bursting regimes. NbO₂ is a transition‑metal oxide that exhibits a sharp, temperature‑driven metal‑insulator transition (MIT) around 800 °C. When an electric field heats the material above this transition, its resistance drops abruptly, producing a high‑current pulse. By exploiting this intrinsic non‑linearity, the authors construct a neuron consisting of two locally active NbO₂ memristors connected in a cross‑coupled configuration. Each memristor is engineered with a distinct threshold voltage (Vth) and thermal capacitance through careful process tuning, allowing them to fire sequentially under a common input stimulus.

The experimental platform applies a DC bias (Vin) combined with rectangular voltage pulses of variable width (τ). By sweeping Vin and τ, the authors map out a two‑dimensional operating space that delineates a spiking region—characterized by isolated, periodic spikes—and a bursting region—where multiple spikes occur in rapid succession within a single burst. The transition boundary is found to be governed by the interplay between the heating dynamics of the first memristor (which switches on when Vin exceeds its Vth) and the thermal coupling to the second memristor. Once the first device turns on, the resulting current surge raises the temperature of the second device, triggering its own transition. The heat diffusion time and the electrical recovery time of the devices compete, producing a finite window during which a cascade of spikes can be generated before the system cools back below the transition threshold.

A key contribution is the quantitative control of the number of spikes per burst (Nspk). The authors demonstrate that Nspk can be tuned from 1 to 7 by adjusting Vin amplitude and τ. Higher Vin accelerates the temperature rise, leading to more rapid successive activations, while longer τ allows more thermal energy to accumulate, extending the burst duration. This controllability enables a novel “spike‑count encoding” scheme: the integer Nspk itself becomes the information carrier, rather than the precise timing or amplitude of individual spikes.

To showcase the computational utility of this encoding, the authors integrate the memristive neurons into a single‑layer perceptron network. Input patterns are mapped to distinct Nspk values, and the output neuron integrates the total spike count during a predefined observation window, applying a thresholded activation function. Unlike conventional voltage‑level encoding, spike‑count encoding eliminates the need for precise analog voltage discrimination and reduces sensitivity to timing jitter, simplifying circuit design and lowering energy consumption. In a classification task involving ten handwritten digit patterns reduced to four output classes, the perceptron achieved over 92 % accuracy. Moreover, when Gaussian noise was added to the input signals, the network retained above‑85 % accuracy, indicating robustness inherent to the burst‑based representation.

The paper also explores information transmission using burst coding. By assigning two bits of data to each possible Nspk value (0–3 spikes), the authors demonstrate simultaneous transmission of four distinct symbols over a single channel. Compared with a baseline voltage‑level modulation scheme, the burst‑based method reduces both energy per bit (by ~40 %) and latency (by ~30 %).

In summary, the study establishes that NbO₂ memristors, with their thermally induced abrupt resistance change, can be harnessed to create compact, low‑power neurons capable of both spiking and bursting. The ability to deterministically set the number of spikes per burst provides a straightforward, noise‑tolerant encoding mechanism that can be directly employed in perceptron‑type networks and simple communication links. The authors conclude that this burst‑control paradigm opens a new design space for hardware spiking neural networks, and they outline future directions such as scaling the memristor fabrication for tighter Vth distribution, synchronizing bursts across large arrays, and integrating learning rules (e.g., spike‑timing‑dependent plasticity) with the spike‑count encoding framework.