Neural Processes
Overview
Neural Processes study how to predict a function from a small set of observed input-output pairs, by learning across many related tasks and returning a distribution of possible outputs. Conditional, attentive, latent-variable, convolutional, and transformer variants differ in how they aggregate context and represent uncertainty; extensions use spectral features or neural operators for periodic data and physical fields. These methods are useful for sparse or irregularly sampled settings such as spatial interpolation, time-series forecasting, and partial-observation PDE prediction. A central tension is balancing expressive context interactions and calibrated uncertainty with consistency, computational efficiency, and the limits imposed by finite representations and amortized inference.
Latest papers
62 papers5 papers in the last four weeks, up 67% on the four weeks before. 0.0% of all new papers.