TL;DR
This paper introduces a novel parallel Convolutional Neural Network (CNN) architecture implemented on memristive crossbar arrays for real-time epileptic seizure detection and prediction. By parallelizing convolution operations and leveraging in-memory computing with RRAM, the system achieves significant reductions in latency and network parameters while maintaining state-of-the-art accuracy, paving the way for more efficient and compact wearable epilepsy monitoring devices.
What Problem Does This Paper Address?
Epileptic seizure detection and prediction are crucial for improving the quality of life for millions affected by this neurological disorder. While advanced algorithms like CNNs have shown great promise, their hardware implementation, especially for real-time, power-constrained applications like wearable devices, remains a significant challenge. Conventional CMOS technology struggles with the high power consumption and latency associated with processing numerous EEG channels. This paper addresses this gap by proposing a novel hardware-efficient architecture that utilizes memristive crossbar arrays for in-memory computing, aiming to drastically reduce latency and power consumption while maintaining high accuracy.
Methods
The core of this work is a novel parallel CNN architecture designed for efficient hardware implementation. Key methodological contributions include:
- Parallel Convolutional Layers: Instead of sequential processing, the authors parallelize the execution of convolution layer kernels across separate analog crossbars. This is a departure from traditional CNN architectures and significantly reduces inference latency.
- In-Memory Computing with RRAM: The CNN is implemented on analog crossbar arrays composed of Resistive Random-Access Memory (RRAM) devices. This "in-memory computing" paradigm allows computations (specifically Vector-Matrix Multiplications) to occur directly within the memory array, minimizing data movement and its associated energy and latency costs.
- Model Search and Selection: A systematic methodology (Algorithm 1) was employed to search for and select an optimized network architecture, balancing performance (accuracy) with hardware constraints (parameter count, tile size).
- Stuck Weight Offsetting: A novel, simplified methodology is proposed to mitigate performance degradation caused by stuck-on ($R_{\text{ON}}$) or stuck-off ($R_{\text{OFF}}$) memristor weights, a common non-ideality in RRAM devices. This method aims to recover accuracy without requiring retraining.
- Quantization-Aware Training (QAT): To address the performance impact of low-resolution Analog-to-Digital Converters (ADCs) and Digital-to-Analog Converters (DACs) in resource-constrained systems, QAT is investigated to mitigate accuracy loss.
Datasets
The authors evaluated their system using three publicly available datasets:
- University of Bonn EEG dataset: Used for epileptic seizure detection.
- CHB-MIT Scalp EEG dataset: Used for epileptic seizure prediction.
- SWEC-ETHZ iEEG dataset: Also used for epileptic seizure prediction.
These datasets are standard benchmarks in the field, allowing for direct comparison with existing methods.
Key Insights and Ideas
- Parallelism is Key for Latency Reduction: The most significant insight is that parallelizing convolution operations across multiple crossbars, rather than processing them sequentially, can lead to a dramatic reduction in latency (orders of magnitude).
- RRAM for Efficient In-Memory Computing: Leveraging RRAM for in-memory computing is crucial for achieving the desired low-power and high-density implementation of CNNs for real-time applications.
- Stuck Weights are a Major Hurdle, but Mitigatable: Stuck memristor weights are a significant challenge for RRAM-based accelerators. The proposed stuck weight offsetting method offers a practical way to recover substantial accuracy without the overhead of retraining.
- Patient Variability Matters: The effectiveness of non-idealities and the proposed mitigation strategies varies significantly across different patients, highlighting the need for personalized or adaptive approaches in future systems.
Key Results
- High Accuracy: The proposed CNN architecture achieved 5-fold cross-validation accuracies of 99.84% for seizure detection and 99.01% and 97.54% for seizure prediction on the respective datasets.
- Reduced Parameters: The architecture requires significantly fewer network parameters, with reductions of 2-2,800x compared to State-Of-The-Art (SOTA) CNNs.
- Latency Reduction: The parallelized memristive implementation achieves approximately 2 orders of magnitude reduction in latency compared to SOTA hybrid Memristive-CMOS DL accelerators.
- Stuck Weight Recovery: The stuck weight offsetting methodology recovered up to 32% of lost accuracy due to stuck $R_{\text{ON}}/R_{\text{OFF}}$ devices.
- Hardware Estimates: The CNN component is estimated to consume approximately 2.791W of power and occupy an area of 31.255 mm² in a 22nm FDSOI CMOS process.
Limitations and Future Work
- Non-Ideality Impact: While the paper investigates the impact of various non-idealities (weight resolution, write error, ADC/DAC resolution, device conductance range, source/line resistance, stuck weights), the system's performance is still affected, especially at higher percentages of stuck devices.
- Stuck Weight Offsetting Limitations: The proposed stuck weight offsetting method cannot fully recover performance if both complementary weights of a bipolar synapse are stuck simultaneously.
- Hardware Simulation: The work is based on simulations, and real-world hardware implementation and testing are necessary for validation.
- Future Work: The authors suggest investigating analog SRAM for further power and area reduction, exploring vertical stacking of crossbars, and implementing partial sensing approaches. They also aim to investigate these aspects in future research.
Summary updated