learning-to-learn-from-data

Code for paper "Learning to learn from data: using deep adversarial learning to construct optimal statistical procedures" by A. Luedtke, M. Carone, N. Simon, and O. Sofrygin. [link]

Environment

All numerical experiments were performed on AWS GPU instances (p2.xlarge and p3.2xlarge), using

• [Julia 0.6.2], and the following Julia packages (versions to the right):
• Knet 0.9.0
• Flux 0.5.1
• CuArrays 0.5.0
• ArgParse 0.5.0
• BSON 0.1.3
• CLArrays 0.1.3
• CLFFT 0.5.2
• CUDAnative 0.6.3
• CURAND 0.0.4
• Clustering 0.9.1
• Distributions 0.15.0
• JLD 0.8.3
• JLD2 0.0.5
• StatsBase 0.21.0

Since Julia version 0.6.2 is no longer maintained, we also provide a publicly available AWS image configured to run our numerical experiments on AWS GPU instances. This image can be accessed by searching for the following AWS Image ID in the region us-east-2a:

• ID: ami-05aa8ea8a7f911839

We note that Example 2 (poorly conditioned point estimation), and Example 2 only, was run in Julia version 0.7.0. Therefore, we have provided a second AWS Image that can be used to run this example:

• ID: ami-0e951811ae6a3a747

Installation

Install the appropriate versions of Julia/Knet/Flux using the following instructions:

git clone git://github.com/JuliaLang/julia.git
cd julia
git checkout v0.6.2
make -j Nproc

The repository also includes several R scripts. All of these scripts were run on a private cluster, and therefore cannot be run on the AWS Images that we have provided without installing the required packages from CRAN.

Replicating numerical experiments

See the specific folders for instructions on running the numerical experiments.

01_pointEstimation_basic

This folder contains instructions and code for running the first 3 point-estimation examples (mean, sd & Neal example with inconsistent MLE).

02_pointEstimation_illPosed

This folder contains instructions and code for running the two poorly-conditioned linear regression problems.

03_prediction_binary

This folder contains instructions and code for running the individual-level prediction problems for binary outcome.

04_prediction_cluster_GMM

This folder contains instructions and code for running the clustering example in two-component Gaussian mixture model.

05_confidence_region_construction

This folder contains instructions and code for running the experiments for confidence region construction.

Citation

If you use our code, please consider citing the following:

Luedtke A, Carone M, Simon N, Sofrygin, O. (2020). Learning to learn from data: using deep adversarial learning to construct optimal procedures. Science Advances.