GPLFR Predicts Alien Climate Maps 4x Faster With Less Data
GPLFR unifies compression and prediction in a single GP objective, cutting training data needs by 4x for high-dimensional scientific emulation tasks
PRO- GPLFR is a new Gaussian process method that jointly learns data compression and prediction in one objective, replacing the standard two-step PCA-GP pipeline.
- On a synthetic benchmark, GPLFR matches PCA-GP accuracy with roughly 4x fewer training examples, a major win in data-scarce scientific settings.
- The method is demonstrated by building the first spatially resolved emulator of rocky exoplanet GCMs, predicting full climate maps from 8 planetary parameters.
- The key innovation is analytically marginalizing decoder weights so the model learns a compression basis optimized for prediction, not reconstruction.
- Limitations include GP's standard cubic scaling with data and an assumption that outputs lie on a linear low-dimensional manifold.
- Code is open-source on GitHub; applicable to any domain with expensive simulations and high-dimensional spatial or spectral outputs.
Running a single 3D general circulation model (GCM) of a rocky exoplanet's atmosphere can take significant computational resources -- roughly 50 cores running in parallel for anywhere from a few days to a few weeks. That means building a full picture of how planetary climate varies across parameters like CO2 levels, stellar flux, or surface gravity requires dozens or hundreds of such runs. The resulting dataset is tiny by machine learning standards, yet each simulation spits out a spatially resolved climate map with thousands of grid points. You have very few examples, each with a very high-dimensional output. This is one of the hardest regimes for any regression model.
A new paper from researchers at the University of Exeter and the University of Colorado proposes a method called Gaussian Process Latent Factor Regression (GPLFR) that is specifically designed for this low-data, high-dimensional-output setting. The authors validate it by building what they describe as the first spatially resolved emulator of global climate models for rocky exoplanets.
The two-step trap
The standard approach to this kind of problem is a two-stage pipeline: first compress the high-dimensional output using PCA (principal component analysis), then fit a separate Gaussian process (GP) to predict each compressed component. Multi-output Gaussian processes excel in low-data regimes but typically struggle with high-dimensional outputs, while compress-then-predict pipelines like PCA-GP handle high dimensionality but rely on bases optimized for reconstruction rather than prediction.
That last point is the crux of the problem. PCA finds the directions that explain the most variance in the training outputs -- but variance in the outputs is not the same as variance that is predictable from the inputs. If the most important input-driven signal lives in a direction that PCA ranks low, it gets discarded before the GP ever sees it. The two stages are decoupled, and there is no way to recover from a bad compression.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
