Predicate Neuron：梯度学习中谓词关系的可微量化
原标题：Predicate Neuron: Differentiable Quantization of Predicate Relations in Gradient-based Learning
AI 摘要
该论文提出 Native Predicate Neuron，一种可微分的逻辑推理神经模块，通过单参数操作选择与定制梯度机制，在训练前向传播中执行硬逻辑运算、反向传播保持可微，从而将一阶逻辑量化原生嵌入神经网络。作者在全连接、卷积、循环和 Transformer 架构及 MNIST、Wikitext 数据集上训练评估，结果显示该模块加速收敛并普遍提升测试准确率、降低损失，在需要处理不连续或单调推理的场景提升尤为显著。模块已以 Keras、PyTorch 自定义层及 Python 包形式开源发布。
正文节选
Insert JAIR AE Name \JAIRTrack Predicate Neuron: Differentiable Quantization of Predicate Relations in Gradient-based Learning Abstract. Background: In machine learning, learned symbolic approaches model reasoning by extracting data patterns that can be expressed as formal logic rules or by imposing weight-based constraints that guide neurons to approximate logical relationships. However, these approaches typically rely on static architectures that prevent them from being widely incorporated int