Towards provably efficient quantum algorithms for large-scale machine learning models

Data for "Towards provably efficient quantum algorithms for large-scale machine-learning models". Error.txt includes the error proxy estimated due to Carleman linearization. The estimates are obtained using Hessian eigenvalues. Hessians.zip contains hessian eigenvalue grids and densities. Most files are for the 7 M parameter mode, and resnet_422-4-* are for the 103 M parameter model. Accuracy.txt contains the sparse training model accuracy on classifying the test set with CIFAR-100, as well as the loss values. Hessian_vrification.ipynb contains the code to generate the supplementary verification of Hessian eigenvalues on the error properties of Carleman linearization plots. The initial conditions are random.本数据集配套于论文《可证明高效的大规模机器学习模型量子算法》（Towards provably efficient quantum algorithms for large-scale machine-learning models）。其中，Error.txt 包含基于卡尔曼线性化（Carleman linearization）得到的误差代理估计值，该估计值通过海森矩阵（Hessian）特征值计算获得。Hessians.zip 存储海森矩阵特征值网格与密度分布数据。绝大多数文件对应7百万参数规模的模型，而以resnet_422-4-*命名的文件则对应103百万参数规模的模型。Accuracy.txt 包含基于CIFAR-100数据集测试集的稀疏训练模型分类准确率，以及对应的损失值。Hessian_vrification.ipynb 包含用于生成卡尔曼线性化误差特性的海森矩阵特征值补充验证绘图代码，实验初始条件均为随机生成。