Abstract
Neuromorphic computing is expected to realize fast and energy-efficient artificial neural networks and address the inherent limitations of von Neumann architectures in dedicated communication applications. To realize this vision, we identify the existing challenges in neuromorphic computing and provide a specific solution from the perspectives of device, circuit, and system. At the device level, we fabricate a metal-oxide-based memristor with high stability, low power, and good scalability, serving as the fundamental component of a neuromorphic computing system. At the circuit level, the basic circuit units and necessary peripheral circuits are designed to realize efficient vector-matrix multiplication and different functions, including nonlinear activation operation, subtraction operation, added operation, and so on. At the system level, a flexible neuromorphic computing system with a hardware-friendly training approach is proposed, which can perform effective communication with good trade-off between accuracy and time consumption. This study is expected to achieve the deep integration of nanotechnology, energy-efficient integrated circuits, and neuromorphic computing systems into communication applications.
Introduction
Traditional computing architectures face a significant "von Neumann bottleneck," where the physical separation of memory and processing units leads to excessive energy consumption and latency, particularly in data-intensive tasks like Natural Language Processing (NLP). Neuromorphic computing offers a brain-inspired alternative by performing in-memory computing. However, current memristor-based systems are often hindered by device instability, architectural rigidity, and a continued reliance on external software for training.
This research addresses these hurdles by presenting a "full-stack" neuromorphic solution. We propose a stable Ag/TiOx nanobelt memristor, a reconfigurable modular circuit design, and a fully autonomous hardware-implemented training approach, enabling a flexible system capable of real-time affective communication (sentiment analysis).
Methods
Device Level
The foundation of the system is an Ag/TiOx nanobelt/Ti memristor fabricated via hydrothermal synthesis. This device demonstrates high cycle-to-cycle stability (verified over 500 cycles) and a high $R_{on}/R_{off}$ ratio of approximately $10^4$. Such characteristics are vital for maintaining precise synaptic weights in a neural network.
Circuit Level
To overcome the rigidity of traditional hardware, we developed a modular "Lego-like" library of basic circuit units. This includes a Linear Matrix Operation Circuit (LMOC) for vector-matrix multiplication and a Nonlinear Activation Circuit (AC). By interconnecting these with peripheral circuits for subtraction and addition, the system can be reconfigured to implement various Long Short-Term Memory (LSTM) architectures, including GRU, Peephole LSTM, and ConvLSTM.
System Level
The system employs a hardware-friendly, six-phase training method. On-chip gradient calculation and weight programming are performed using 50ns pulses. This autonomy eliminates the communication overhead typically required when interfacing with external software-based trainers.
Results and Discussion
The proposed system was evaluated using standard sentiment analysis benchmarks: IMDB, SST-2, and Yelp-2 datasets, as well as synthetic data generated via Easy Data Augmentation (EDA). The hardware-based approach demonstrated exceptional efficiency and performance compared to both existing circuit-based and software-based methods.
Key results include:
- Energy Efficiency: The system operates with a power footprint of only 100.99 μW and an area of 74.83 μm², representing a significant reduction compared to previous memristor-based LSTMs.
- Processing Speed: While maintaining accuracy within a 5% margin of software models, the hardware system achieved processing speeds 30 to 35 times faster than software-based implementations.
- Robustness: The system consistently outperformed other state-of-the-art circuit-based methods across Accuracy, F1-Score, and Recall metrics for all tested datasets.
Conclusion
This work demonstrates a viable path toward fully integrated, energy-efficient neuromorphic systems for communication. Future research will focus on utilizing more sustainable, eco-friendly materials for memristor fabrication and developing advanced brain-inspired circuits for real-time online learning in dynamic environments.
Summary updated