SpikeLogBERT uses spiking transformers for energy-efficient log parsing

Researchers present SpikeLogBERT, a spiking neural network framework designed for energy-efficient log parsing that transforms raw system logs into structured event templates.

• The model integrates a spiking transformer architecture with knowledge distillation from a BERT teacher model to preserve semantic representation capability.
• It leverages sparse spike activations and event-driven processing to significantly reduce the number of active operations during inference.
• Experiments on the HDFS dataset show SpikeLogBERT achieves a parsing accuracy of 0.99997, outperforming ANN-based neural log parsing models.
• The approach reduces estimated theoretical energy consumption by up to 62.6% under standard 45nm CMOS assumptions.

This framework offers a method for automated log analysis that maintains high accuracy while substantially lowering the computational cost and energy consumption associated with dense matrix multiplications.