Normalizing Flows (NFs)是一个生成模型系列,具有可操作的分布,其采样和密度评估都是有效和精确的。
被探索的大部分Flows是三角流triangular flows(coupling耦合或autoregressive自回归架构),Residual networks和Neural ODEs也正在积极研究和应用。
NORMALIZING FLOWS
Coupling and Autoregressive Layers | |
Affine Coupling | |
Monotone Functions | |
Autoregressive Flows | |
Probability Density Distillation | |
Convolutional | |
Residual Flows | |
Matrix Determinant Lemma | |
Lipschitz Constrained | |
Surjective and Stochastic Layers | |
Discrete Flows | |
Continuous Time Flows | |
Regularising Trajectories |
NFs研究方向
Inductive biases (归纳性偏置) | |
role of the base measure (基准测量的作用) | |
Form of diffeomorphisms (微分同胚的形式) | |
loss function |
|
Generalisation to non-Euclidean spaces(非欧几里得空间的泛化) | |
flows on manifolds |
|
discrete distributions (离散分布) 去量化dequantization,(即在离散数据中加入噪声,使其成为连续数据) |
|