Scholarlink

This paper introduces Differentiable Optimal Discrete Value Quantization (DODVQ), a training method for Spiking Neural Networks (SNNs) for inference on neuromorphic hardware with strict synaptic weight constraints. Unlike conventional approaches that derive a discrete set of weights by applying clustering algorithms (e.g., k-Means) to full-precision or QAT-trained weights, DODVQ treats the discrete values themselves as learnable parameters co-optimized with the model. We demonstrate that for models constrained to only four unique weight values ( \(M=4\) ), our DODVQ-SNN drastically outperforms a baseline SNN created with a conventional QAT-plus-clustering pipeline. On the FashionMNIST and EEG Eye State datasets, DODVQ recovers nearly all the performance of full-precision models, whereas the baseline method suffers a catastrophic drop in accuracy. Our work provides a robust pathway for developing high-performance SNNs that are directly compatible with the hardware constraints of energy-efficient neuromorphic processors.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Differentiable Optimal Discrete Value Quantization for Spiking Neural Networks on Weight-Constrained Neuromorphic Hardware

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ivan Kartashov,
  • Mariia Pushkareva,
  • Yakov Karandashev

摘要

This paper introduces Differentiable Optimal Discrete Value Quantization (DODVQ), a training method for Spiking Neural Networks (SNNs) for inference on neuromorphic hardware with strict synaptic weight constraints. Unlike conventional approaches that derive a discrete set of weights by applying clustering algorithms (e.g., k-Means) to full-precision or QAT-trained weights, DODVQ treats the discrete values themselves as learnable parameters co-optimized with the model. We demonstrate that for models constrained to only four unique weight values ( \(M=4\) ), our DODVQ-SNN drastically outperforms a baseline SNN created with a conventional QAT-plus-clustering pipeline. On the FashionMNIST and EEG Eye State datasets, DODVQ recovers nearly all the performance of full-precision models, whereas the baseline method suffers a catastrophic drop in accuracy. Our work provides a robust pathway for developing high-performance SNNs that are directly compatible with the hardware constraints of energy-efficient neuromorphic processors.