Reveals that 'learned priors' in inverse problems often behave as simple lookup tables that memorize training data rather than learning distributions.
March 23, 2026
Original Paper
On the role of memorization in learned priors for geophysical inverse problems
arXiv · 2603.19629
The Takeaway
Crucial for practitioners using diffusion-based priors in scientific imaging or geophysics. It suggests that your 'state-of-the-art' reconstruction might just be a reweighted average of your training samples, limiting its ability to generalize to novel physical phenomena.
From the abstract
Learned priors based on deep generative models offer data-driven regularization for seismic inversion, but training them requires a dataset of representative subsurface models -- a resource that is inherently scarce in geoscience applications. Since the training objective of most generative models can be cast as maximum likelihood on a finite dataset, any such model risks converging to the empirical distribution -- effectively memorizing the training examples rather than learning the underlying