Denoising Neural Processes model noisy inputs during meta-learning
Original title: DnNP: Denoising Input Uncertainty in Neural Processes
On real air-quality sensors, DnNP cuts PM2.5 RMSE by up to 30% and NLL by up to 73% versus standard neural processes.
The paper introduces Denoising Neural Processes, an extension of neural processes that models input noise instead of assuming noise-free inputs. DnNPs are trained with variational inference and jointly denoise corrupted inputs while meta-learning. On synthetic and real-world data the authors report better predictions and uncertainty estimates than standard neural processes. They also argue the Bayes-optimal predictor under noisy inputs lies in the DnNP class, and that fixed-variance plug-in models leave an irreducible KL gap.
- Standard neural processes assume noise-free inputs.
- On PM2.5 interpolation, RMSE falls by up to 30% versus standard NPs.
- Negative log-likelihood falls by up to 73% on the same task.
- Training uses variational inference on synthetic and real-world data.
From the original article
1 min read at arxiv.orgNeural Processes (NPs) are flexible meta-learning models capable of uncertainty quantification across diverse applications. However, the standard NP framework assumes access to noise-free inputs. This is unrealistic in many practical settings, where sensor noise or measurement errors are unavoidable.
By Fatemeh Tohidian · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.