.png){kind=icon}
.png){kind=icon}
Return
Post-training quantization for efficient ANN-SNN conversion
.png){kind=icon}
PRE
AI
DOI：10.1016/j.neunet.2025.107832.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
• We leverage Post-Training Quantization (PTQ) to efficiently convert pre-trained ANNs into SNNs. PTQ, a widely adopted technique in quantizing ANNs, allows us to swiftly transform an ANN into an SNN with minimal calibration. Our experiments demonstrate that this approach enables us to achieve higher-precision SNN models with fewer samples and a shorter calibration time. • We theoretically investigate the sources of conversion error during the ANN-SNN conversion process. To mitigate these errors, we propose a channel-wise thresholding scheme by reparameterizing the affine scaling factor γ in the Batch Normalization layer, enabling channel-level thresholds to approximate the effectiveness of sample-wise thresholding. Additionally, we introduce Householder reflection thresholding as a complementary technique to further minimize conversion errors. • Our experimental results on both CNN-based and Transformer-based networks, using static image and neuromorphic datasets, demonstrate the effectiveness of our proposed method in improving classification accuracy.
Keywords:
Post-Training Quantization
Spiking Neural Networks
Batch Normalization
Channel-wise Thresholding
Householder Reflection