Abstract
Mental health issues, including depression and chronic stress, are increasingly prevalent social concerns. Facial expression recognition serves as a vital tool for identifying emotional states, yet current automated systems often suffer from high computational costs and low efficiency. This paper presents a memristive circuit design of a Sequencer network for human emotion classification. By utilizing eco-friendly two-dimensional (2D) materials to fabricate memristors, we develop a hardware-deployable Sequencer block centered on a bidirectional long short-term memory (BiLSTM) circuit. Experimental results indicate that this memristive hardware approach achieves computational efficiency and cost-effectiveness comparable to existing software-based methods, offering a sustainable path for real-time mental health monitoring.
Introduction
Automatic human emotion classification is essential for early mental health intervention. However, modern deep learning models, particularly those based on the Transformer architecture, are computationally expensive due to the "von Neumann bottleneck"—the constant data movement between processors and memory. This leads to high energy consumption and latency, making deployment on edge devices difficult.
To address these challenges, this research moves toward neuromorphic computing using memristors. These components can simultaneously store data and perform calculations, enabling a low-cost, high-efficiency hardware implementation of the "Sequencer" network. The Sequencer model mimics the spatial processing capabilities of Vision Transformers but replaces complex self-attention mechanisms with more hardware-friendly components.
Methods and Architecture
The proposed system integrates material science with advanced circuit architecture through a multi-step process:
1. Eco-Friendly Memristor Fabrication
The team fabricated specialized memristors using MoSe₂-doped Selenium (Se) microwires and Ag-Au electrodes. This 2D-material-based device acts as the "synapse" of the network. Testing confirmed high device-to-device and cycle-to-cycle stability over 500 cycles, ensuring the neural network weights remain stable during operation.
2. The Sequencer Network Design
Unlike standard Vision Transformers (ViT) that rely on mathematically complex self-attention layers, the Sequencer network utilizes Bidirectional Long Short-Term Memory (BiLSTM) blocks. BiLSTMs are highly effective at processing sequences (such as rows and columns of pixels) and are significantly easier to implement in physical circuitry.
3. Circuit Implementation
The authors designed specific memristive modules for every network component, including BiLSTM cells using memristor crossbars for linear operations and specialized circuits for nonlinear functions (tanh and sigmoid). All modules utilize voltage-mode cascading, allowing different stages—such as Layer Normalization, ReLU activation, and Global Average Pooling—to be connected seamlessly.
Experimental Results
The memristive Sequencer network was validated using two standard benchmarks: the CK+ (Extended Cohn-Kanade) dataset and the FER-2013 "in-the-wild" dataset.
- Efficiency Gains: The hardware-based system achieved a massive reduction in computational overhead, with a mean processing time of 125.4 seconds. In contrast, software-based versions of similar networks required between 2013.7 and 2335.9 seconds, representing a nearly 16x speedup.
- Accuracy Trade-offs: The system achieved an accuracy of 92.3% on CK+ and 72.3% on FER-2013. While these figures are approximately 3-4% lower than top-tier software models, the trade-off is justified by the significant gains in speed and energy efficiency.
Conclusion and Future Work
This work demonstrates that memristive hardware can effectively implement complex neural architectures like the Sequencer network for real-world tasks. While a small accuracy gap remains compared to high-precision software, future refinements in circuit noise handling and device variability could close this margin. Additionally, moving from off-chip training in MATLAB to fully autonomous on-chip learning will be a primary focus for future iterations of this technology.
Summary updated