DnNP: Denoising Input Uncertainty in Neural Processes
DnNP cut prediction error by up to 30% on air-quality sensor data by accounting for noisy inputs.
Standard Neural Processes assume noise-free inputs, which can undermine predictions and uncertainty estimates when measurements are corrupted. DnNP jointly learns to denoise inputs and make predictions, using variational inference. In PM2.5 spatial interpolation, the authors report reductions of up to 30% in RMSE and 73% in negative log-likelihood relative to standard Neural Processes. They also show theoretically that DnNP can represent the Bayes-optimal predictor under noisy inputs.
ArXiv · AI/CL/LG's note
Standard Neural Processes assume noise-free inputs, which can undermine predictions and uncertainty estimates when measurements are corrupted. DnNP jointly learns to denoise inputs and make predictions, using variational inference. In PM2.5 spatial interpolation, the authors report reductions of up to 30% in RMSE and 73% in negative log-likelihood relative to standard Neural Processes. They also show theoretically that DnNP can represent the Bayes-optimal predictor under noisy inputs.
ArXiv · AI/CL/LG's note
score 4