Metastable Learning

This has been tested on Julia 1.9.2. The checked-in manifest was resolved with Julia 1.9.3, so use Julia 1.9.x for the most reliable reproduction.

Code to recreate the numerical experiments in the paper is in the Experiments folder. The root Project.toml and Manifest.toml contain the Julia environment with the packages needed to run the files.

Julia environment

From the repository root, start Julia with the project environment active:

julia --project=.

The first time you use the repository, instantiate the environment from the Julia REPL:

using Pkg
Pkg.instantiate()

For one-off scripts, pass the environment on the command line:

julia --project=. path/to/script.jl

Curie-Weiss experiments

The main Curie-Weiss experiments are in Experiments/CW_tests. The scripts in this folder use relative paths, so run them from inside that directory:

cd Experiments/CW_tests

Collect samples and learn the Curie-Weiss model parameters:

julia --project=../.. cw_datacollect.jl

This writes or updates data_cw.jld2. The defaults match the plotting scripts: n_spins = 5000, J * n_spins = 1.2, h = 0.04, Glauber sampling, and M = 2^22, 2^24, ..., 2^32. These are large runs; edit the parameter lists near the top of cw_datacollect.jl for smaller smoke tests.

Create the summary plots from data_cw.jld2:

julia --project=../.. cw_plotter.jl

This writes error_cw.pdf, prob_cw.pdf, and mag_cw.pdf.

Create the free-energy histogram plot:

julia --project=../.. free_energy_plotter.jl

This writes FE_histogram.pdf. If the cached sample file named in the script is not present, the script will generate new Glauber samples, which can be slow.

Spin-glass learning experiments

Spin-glass learning experiments are in Experiments/PottsSGLearning/n12_beta_sweep. These scripts use @__DIR__ to find the repository paths, so they can be run from the repository root.

To reproduce the checked-in plots from the checked-in result files:

julia --project=. Experiments/PottsSGLearning/n12_beta_sweep/plotter.jl
julia --project=. Experiments/PottsSGLearning/n12_beta_sweep/plotter_energy.jl
julia --project=. Experiments/PottsSGLearning/n12_beta_sweep/plotter_panelb.jl

The scripts write:

• n12_24_beta_sweep.pdf from results.jld2 and results_n24.jld2
• n12_24_energy_sweep.pdf from results.jld2 and results_n24.jld2
• n12_24_panelb.pdf from results_panelb.jld2

To regenerate the result files before plotting, run:

julia --project=. Experiments/PottsSGLearning/n12_beta_sweep/datacollector.jl
julia --project=. Experiments/PottsSGLearning/n12_beta_sweep/datacollector_n24.jl
julia --project=. Experiments/PottsSGLearning/n12_beta_sweep/datacollector_panelb.jl

The Potts utilities used by these experiments live in PottsPSpin, and the Potts pseudolikelihood learner is in Experiments/PottsSGLearning/potts_sumoflocal_PLE.jl.

License

This code is provided under a BSD license as part of the Optimization, Inference and Learning for Advanced Networks project, C18014.