Evidential Conditional Neural Processes

The Conditional Neural Process (CNP) family of models offer a promising direction to tackle few-shot problems by achieving better scalability and competitive predictive performance. However, the current CNP only capture overall uncertainty for prediction made on target data point. They lack systematic fine-grained quantification distinct sources that are essential model training decision-making under setting. We propose Evidential Processes (ECNP), which replace standard Gaussian distribution used with much richer hierarchical Bayesian structure through evidential learning achieve epistemic-aleatoric decomposition. also leads theoretically justified robustness over noisy tasks. Theoretical analysis proposed ECNP establishes relationship while offering deeper insights roles parameters. Extensive experiments conducted both synthetic real-world demonstrate effectiveness our in various settings.