Conditional F\"ollmer Flow

Conditional F\"ollmer flows are a class of deep generative models using ordinary differential equations (ODEs) to learn complex conditional probability distributions. Current research focuses on improving their efficiency and accuracy for various applications, employing architectures like invertible triangular attention layers and neural network-based velocity field estimations within the ODE framework. These models offer advantages in probabilistic forecasting, particularly for irregularly sampled time series and high-dimensional data, enabling improved uncertainty quantification and out-of-distribution detection across diverse fields like healthcare and climate science. Their ability to learn complex dependencies and provide tractable likelihood calculations makes them a valuable tool for various probabilistic modeling tasks.