Generating populations with GANs: the generator and discriminator

Japanese labels. Resident-card values are fictional, not measured results.

GAN means Generative Adversarial Network, introduced by Goodfellow and colleagues in 2014.[1]

What is learned?

Given survey records of resident attributes, a generator G takes random inputs and creates synthetic records. A discriminator D learns to distinguish observed survey records from generated records. G is then updated to produce records D finds harder to distinguish. Training alternates these updates.

One analogy is a person making practice questions alongside someone judging whether they resemble an existing collection. Both improve, but D is not a judge with knowledge of the entire real population.

A basic VAE uses encoding, decoding, reconstruction, and KL regularization. A basic GAN uses competition between G and D as its training signal. Hybrid methods exist, so this is not an absolute classification of every implementation.

A population example

Suppose fictional source proportions are younger/employed 40%, younger/not employed 20%, older/employed 10%, and older/not employed 30%.

If G generates only younger employed people, D has evidence of a difference from the source. But successfully confusing D does not guarantee minority coverage or real-world accuracy. Check mode collapse—generating mostly a few kinds—and evaluate against independent evidence.

After training, generation follows random input → trained G → attributes. The basic procedure does not require D to rescore every generated record.

Reading code

In vdarsel/PopulationSynthesis, models/GAN/WGAN/model_WGAN_alone.py defines Generator_Tabular, Discriminator_Tabular, and a gradient penalty. train_WGAN_alone.py uses separate optimizers for D and G. During D’s update, fake_data.detach() prevents simultaneously updating G through those generated values.[2]

This is a WGAN-family implementation with a different objective from a standard GAN. Do not apply the elementary description “outputs a probability of being real” to every WGAN. The retrieved code also ends D with Sigmoid; we should inspect that difference rather than assume an exact textbook WGAN-GP implementation. Code availability and reuse permission are separate questions.

What we checked

We read generator/discriminator code statically; we did not train a GAN. Teaching calculations check how a standard GAN generator’s loss changes with D’s output and how even a copy-based generator can distort a distribution. These are not real-population performance measurements or reproduction of the authors’ experiments.

The adversarial training signal and mode collapse are useful comparison points when studying diffusion models’ denoising approach.

1. Goodfellow et al. 2014, method section 3.
2. Public code read statically.