Input vector for the generator
Remember GAN
Full name: Generative Adversarial Network
Memory aid: Generator And critic Negotiate as the memory cue.
Adversarial training can be unstable and does not guarantee diverse high-quality outputs.
Study area: AI & machine learning · Batch 9
Relevant study for: Azure AI Fundamentals AI-901 · Azure AI Apps and Agents Developer AI-103 · Azure AI Fundamentals (retired) AI-900 · Azure AI Engineer (retired) AI-102
Practice with exam flashcards →AI Models & Assistants / Knowledge library
Generative Adversarial Networks (GANs)
Generative models trained through a generator-discriminator game
Why learn Generative Adversarial Networks (GANs)?
GANs can produce sharp samples but training may be unstable and can miss modes.
Where it came from
GAN training pits a generator against a discriminator. The generator learns to produce samples that the discriminator cannot distinguish from training data.
Build your understanding
Work through these exercises in order. Try the task before opening a reference.
- Identify GAN roles: What does the GAN discriminator learn?
- State the generator objective: What pressure trains the generator?
- Recognize mode collapse: A GAN generates many high-quality but nearly identical faces. What failure is this?
- Diagnose training balance: Why can an overwhelmingly strong discriminator stall early training?
- Evaluate generators: Why is visual inspection alone insufficient?
Documentation and sources
- Primary reference (arxiv.org)
Connect the ideas
Explore next
Real World Example
Synthetic Architectural Blueprint Generation
Architecture and urban planning firms requiring rapid iteration of floor plan layouts.
The business problem
Manual drafting of diverse floor plan variations is time-consuming and often constrained by the architect's habitual design patterns, leading to repetitive layouts.
How this topic solves it
By employing Generative Adversarial Networks, firms can train a generator to produce novel floor plan layouts while a discriminator learns to distinguish between authentic architectural designs and synthetic ones, forcing the generator to produce increasingly realistic and varied outputs.
- Curate a dataset of existing, high-quality architectural floor plan images.
- Initialize a generator network to synthesize floor plan pixels from random noise.
- Initialize a discriminator network to classify inputs as either real architectural data or generator-produced samples.
- Execute a minimax training loop where the generator maximizes the discriminator's classification error.
- Iteratively refine the generator until it produces layouts that are indistinguishable from professional designs.
The flow shows the adversarial interaction between the generator and discriminator.
Creates synthetic floor plans
Actual architectural blueprints
Classifies real vs synthetic
- Latent Noise → transforms → Generator
- Generator → presents synthetic → Discriminator
- Real Data → presents ground truth → Discriminator
- Discriminator → provides feedback → Generator
Expected business outcome
Architects receive a diverse set of high-fidelity, plausible floor plan concepts that serve as creative starting points for final drafting.
Limitations and checks
Mode collapse, where the generator produces a limited variety of outputs despite the training data diversity; this is checked by monitoring the visual variance of generated samples over training epochs.