论文标题
使用非均匀数据的神经网络训练中的调谐频率偏差
Tuning Frequency Bias in Neural Network Training with Nonuniform Data
论文作者
论文摘要
过度参数化的神经网络（NNS）的小概括误差可以通过频率偏见现象来部分解释，在频率偏置现象中，基于梯度的算法最大程度地减少了低频失误，然后再降低高频残差。使用神经切线内核（NTK），可以为训练提供理论上严格的分析，其中数据是从恒定或分段构剂概率密度绘制的数据。由于大多数训练数据集不是从此类分布中汲取的，因此我们使用NTK模型和数据依赖性正规规则来理论上量化了完全不均匀的数据的NN训练的频率偏差。通过用精心选择的Sobolev规范替换损失函数，我们可以进一步扩大，抑制，平衡或逆转NN训练中的内在频率偏置。
Small generalization errors of over-parameterized neural networks (NNs) can be partially explained by the frequency biasing phenomenon, where gradient-based algorithms minimize the low-frequency misfit before reducing the high-frequency residuals. Using the Neural Tangent Kernel (NTK), one can provide a theoretically rigorous analysis for training where data are drawn from constant or piecewise-constant probability densities. Since most training data sets are not drawn from such distributions, we use the NTK model and a data-dependent quadrature rule to theoretically quantify the frequency biasing of NN training given fully nonuniform data. By replacing the loss function with a carefully selected Sobolev norm, we can further amplify, dampen, counterbalance, or reverse the intrinsic frequency biasing in NN training.
