Core Concepts
Autograd can be used to take the derivative of neural network models of cumulative distribution functions (cdfs) to obtain approximations of the corresponding probability density functions (pdfs) of test statistics, which is useful for simulation-based frequentist inference.
Abstract
The authors explore the use of autograd to take the derivative of neural network models of cdfs in order to obtain approximations of the corresponding pdfs of test statistics. This is motivated by the fact that simulation-based frequentist inference methods require accurate modeling of either the p-value function or the cdf of the test statistic.
The paper first considers the classic ON/OFF problem in astronomy and high-energy physics as a benchmark example. It is found that the ALFFI algorithm, which models the cdf as the mean of a certain discrete random variable, does not yield a sufficiently accurate smooth model of the cdf, and hence the derived pdf exhibits sharp fluctuations.
The authors then directly model the empirical cdf as a function of the model parameters and the test statistic, which yields much better results. Conformal inference is used to quantify the uncertainty in the cdf and pdf models.
The insights gained from the ON/OFF example are then applied to the SIR model in epidemiology, which exemplifies the utility of the methods for inference with intractable statistical models. Various techniques for uncertainty quantification, including Bayesian neural networks and bootstrap, are explored and compared.
The paper concludes that directly modeling the empirical cdf is a viable approach, and that conformal inference provides a simple benchmark for calibrating other uncertainty quantification methods.
Stats
The authors generate synthetic data for the ON/OFF problem and the SIR model to train and evaluate their neural network models.
Quotes
"Autograd does not use finite difference approximations."
"Strictly speaking, Eq. (1) applies only if x is from a continuous set. However, discrete distributions are frequently encountered in high-energy physics and other fields, and are often approximated by continuous distributions through suitable coarse-graining of x."
"The key issue is whether a sufficiently accurate model of the cdf F(x | θ) can be constructed."