Memory-Efficient Spiking Neural Networks Using MLC FeFETs and Mixed Quantization

This paper proposes a memory-efficient approach for Spiking Neural Networks (SNNs) using mixed quantization and Multi-Level Cell (MLC) FeFETs in in-memory computing (IMC) architectures. By assigning higher precision to more sensitive layers, we minimize memory usage without compromising accuracy. We further optimize memory through a mixed MLC/SLC configuration, storing less significant bits in MLCs and critical bits in SLCs, reducing the memory footprint while maintaining network performance. Additionally, error-aware training techniques, including Fault-Aware Weight Regularization, are employed to improve robustness against MLC-induced errors. Experimental results demonstrate substantial memory reductions, improved error resilience, and high accuracy retention, making this approach well-suited for energy-efficient edge AI applications.