SelectPFN
A model based on a prior-data fitted network (PFN) that jointly performs Bayesian variable selection and sparse nonlinear regression prediction. It outputs predictions and the posterior inclusion probability of each variable in a single forward pass, and achieves exact finite-sample false discovery rate control.
Timeline (last 90 days)
SelectPFN is a model based on prior-data fitted networks (PFN) that can jointly perform Bayesian variable selection and prediction and is designed specifically for sparse nonlinear regression tasks; in a single forward pass it not only gives prediction results but also returns the posterior inclusion probability of each variable.
SelectPFN's predictions are computed based only on a sampled subset of variables, and the excluded variables can be mathematically proven to be unable to affect the prediction results.
SelectPFN is pretrained in a one-off manner on synthetic tasks through a meta-prior built around a group spike-and-slab sparse input neural network.
SelectPFN can complete a prediction in about 0.04 seconds, and the fundamental reason lies in the amortized inference nature of PFN, which amortizes the per-dataset inference cost into the pretraining stage.
The SelectPFN research team developed Bayesian regret bounds for amortized selection, quantifying the gap between amortized inference and ideal Bayesian inference for the amortized median probability model.
Related research on SelectPFN characterizes that the amount of information needed to detect a variable grows with the prior weight ω of the structure corresponding to that variable, at a rate of approximately (2/λ)log(1/ω), where λ≤1 is an efficiency factor.
Research on SelectPFN proved a negative result: training in the average case does not automatically translate into frequentist guarantees under a fixed ground truth, so controlling the false discovery rate on any given dataset requires additional mechanisms.
SelectPFN generates knockoff statistics through column-permutation equivariance combined with random interleaving of original columns and knockoff columns, achieving exact finite-sample false discovery rate control, and this guarantee holds even when the meta-prior is misspecified.
On nonlinear designs beyond the prior range with p up to 500, SelectPFN's area under the precision-recall curve (AUPRC) improved by 0.23-0.28 compared with lasso, BART, DART, LassoNet, and group-penalized sparse neural networks.
On nonlinear designs beyond the prior range with p up to 500, SelectPFN's false discovery proportion was controlled at about 0.07.
2 more timeline events
All Facts (12)
SelectPFN is a model based on prior-data fitted networks (PFN) that can jointly perform Bayesian variable selection and prediction and is designed specifically for sparse nonlinear regression tasks; in a single forward pass it not only gives prediction results but also returns the posterior inclusion probability of each variable.
50%UnverifiedSelectPFN's predictions are computed based only on a sampled subset of variables, and the excluded variables can be mathematically proven to be unable to affect the prediction results.
50%UnverifiedSelectPFN is pretrained in a one-off manner on synthetic tasks through a meta-prior built around a group spike-and-slab sparse input neural network.
50%UnverifiedSelectPFN can complete a prediction in about 0.04 seconds, and the fundamental reason lies in the amortized inference nature of PFN, which amortizes the per-dataset inference cost into the pretraining stage.
50%UnverifiedThe SelectPFN research team developed Bayesian regret bounds for amortized selection, quantifying the gap between amortized inference and ideal Bayesian inference for the amortized median probability model.
50%UnverifiedRelated research on SelectPFN characterizes that the amount of information needed to detect a variable grows with the prior weight ω of the structure corresponding to that variable, at a rate of approximately (2/λ)log(1/ω), where λ≤1 is an efficiency factor.
50%UnverifiedResearch on SelectPFN proved a negative result: training in the average case does not automatically translate into frequentist guarantees under a fixed ground truth, so controlling the false discovery rate on any given dataset requires additional mechanisms.
50%UnverifiedSelectPFN generates knockoff statistics through column-permutation equivariance combined with random interleaving of original columns and knockoff columns, achieving exact finite-sample false discovery rate control, and this guarantee holds even when the meta-prior is misspecified.
50%UnverifiedOn nonlinear designs beyond the prior range with p up to 500, SelectPFN's area under the precision-recall curve (AUPRC) improved by 0.23-0.28 compared with lasso, BART, DART, LassoNet, and group-penalized sparse neural networks.
50%UnverifiedOn nonlinear designs beyond the prior range with p up to 500, SelectPFN's false discovery proportion was controlled at about 0.07.
50%UnverifiedOn nonlinear designs beyond the prior range with p up to 500, SelectPFN's processing time was only about 0.04 seconds per dataset.
50%UnverifiedSelectPFN's advantages are likewise maintained on real covariate matrices, and the detection rate of variables with only interaction effects rises monotonically as their prior weight increases.
50%