Adaptive LIF-Based Output Neuron Design for Improved Accuracy in Integer-Only Digital Spiking Neural Networks

초록

This work presents a fully digital, fixed-weight spiking neural network (SNN) architecture for deterministic image classification. The system achieves high-precision classification through a systematic four-stage optimization process: multi-feature temporal decoding, adaptive threshold tuning, neuron-specific scaling, and ensemble-based voting. Output neurons integrate leaky membrane potentials and adaptive thresholds, while temporal evidence is encoded using spike counts and first-spike timing via ROM-stored integer coefficients. To improve hardware efficiency and enable high-speed inference, the architecture incorporates data-dependent multiplication skipping and a resource-shared spatial accumulation mechanism. Furthermore, critical hardware bottlenecks are resolved using 2D array unrolling and a two-stage pipelined decision tree. This eliminates routing congestion and combinational delays, achieving a stable 100 MHz operating frequency. All computations are performed using fixed integer arithmetic without additional training, ensuring deterministic behavior. RTL-level evaluation demonstrates that the proposed design improves classification accuracy from 74.5% to 98.7% on the full MNIST test set (10,000 samples). The architecture also exhibits strong generalization, achieving 87.4% accuracy on the complex Fashion-MNIST dataset (10,000 samples) without parameter retuning. These results indicate that carefully engineered temporal dynamics combined with high-speed, integer-only processing substantially enhance inference throughput and energy efficiency in compact digital neuromorphic processors. © 2013 IEEE.

키워드