Neural Bootstrapping Attention for Neural Processes
Neural Processes (NP) learn to fit a broad class of stochastic processes with neural networks. Modeling functional uncertainty is an important aspect of learning stochastic processes. Recently, Bootstrapping (Attentive) Neural Processes (B(A)NP) propose a bootstrap method to capture the functional uncertainty which can replace the latent variable in (Attentive) Neural Processes ((A)NP), thus overcoming the limitations of Gaussian assumption on the latent variable. However, B(A)NP conduct bootstrapping in a non-parallelizable and memory-inefficient way and fail to capture diverse patterns in the stochastic processes. Furthermore, we found that ANP and BANP both tend to overfit in some cases. To resolve these problems, we propose an efficient and easy-to-implement approach, Neural Bootstrapping Attentive Neural Processes (NeuBANP). NeuBANP learns to generate the bootstrap distribution of random functions by injecting multiple random weights into the encoder and the loss function. We evaluate our models in benchmark experiments including Bayesian optimization and contextual multi-armed bandit. NeuBANP achieves state-of-the-art performance in both of the sequential decision-making tasks, and this empirically shows that our method greatly improves the quality of functional uncertainty modeling.
PDF Abstract
