Adversarial learning: how G and D interact in a GAN

What's actually being asked

The basic idea of adversarial learning, how G turns noise into fake samples, how D scores real vs fake, and a workflow diagram tying it together.

Answer

The basic concept

Adversarial learning sets up two models with opposite incentives and trains them simultaneously, using each other's performance as the training signal instead of a fixed labelled dataset telling one model what's right. In a GAN specifically:

• Generator G is the "forger" — it tries to produce samples indistinguishable from real data.
• Discriminator D is the "detective" — it tries to catch which samples are forgeries.

As training progresses, both get better in lockstep: a sharper detective forces the forger to improve, and a better forger forces the detective to look for subtler tells. This push-pull is what "adversarial" refers to — not a cooperative shared loss, but a genuine competition.

Generator: noise to synthetic sample

G is a differentiable network (transposed-convolution stack, or fully connected layers for simpler data) that takes a latent noise vector z, sampled from a simple, fixed distribution p_z — typically z ~ N(0,I) or z ~ U(−1,1) — and maps it to the data space: x_fake = G(z). There's no real input here at all; z supplies all the randomness/variety, and G's learned weights supply the mapping from "arbitrary point in noise space" to "looks like a plausible data sample." Different z's should ideally map to different, diverse, realistic outputs.

Discriminator: scoring real vs generated

D is a binary classifier network — convolutional layers for images, ending in a single sigmoid output D(·) ∈ (0,1), interpreted as "probability this input is real." It's shown both real samples x (from p_data) and fake samples G(z), and trained with ordinary classification loss to push D(x)→1 and D(G(z))→0.

The overall workflow

See the diagram: noise z goes into G, producing a synthetic sample; that synthetic sample, alongside a real sample, goes into D; D outputs a real/fake score; the error in that score is backpropagated two ways — into D's own weights (to sharpen its detection) and, through D, into G's weights (to make its forgeries fool D better next time). The two updates alternate, round after round, rather than happening jointly on one combined loss.

Objective function underlying this workflow

min_G max_D V(D,G) = 𝔼_{x∼p_data}[log D(x)] + 𝔼_{z∼p_z}[log(1 − D(G(z)))] D maximizes log D(x) + log(1−D(G(z))); G, in practice, maximizes the non-saturating log D(G(z)) for healthier gradients rather than literally minimizing log(1−D(G(z))).

Exam tip

If the question specifically says "basic concept," don't dive straight into the formula — spend a sentence or two on the forger/detective intuition first. Examiners awarding "concept" marks are checking you understand why it's called adversarial, not just that you memorized the equation.