ScioVirtualAI 101

AI 101 · Final Course Challenge

Generative AI Lab

Train two competing neural networks to recreate an unknown sensor distribution—then use the evidence to explain what your model learned.

Start the challenge

01 · Configure

Configure your experiment

Train against four fixed signal clusters. Each point is one thermal and radiation sensor reading.

Four Signal Sources

The seed fixes the real dataset and model noise. Small differences may still appear across devices and browser backends.

02 · Predict

Make a hypothesis

Intentional experiments teach you more than random settings.

03 · Observe

Watch the networks compete

Blue circles are fixed real observations. Coral diamonds are generator predictions.

Real Sensor Data

Four Signal Sources · 500 fixed observations

Generated Sensor Data

Live output from random latent noise

Generator output will appear when training begins.

Discriminator Confidence

Warmer regions are judged more likely to be real.

Likely fakeLikely real

04 · Train

Training controls

Ready
Step 0 of 2500%
0:00
Elapsed—
Backend

Live metrics

Read the competition

Why don’t the losses go to zero?

GAN losses often fluctuate because each network changes the other network’s challenge. Coverage and similarity tell you more about the distribution match.

Training diagnosis

Choose settings and train the networks to begin your diagnosis.

Diagnosis uses coverage, occupied regions, generated variance, multimodal coverage, and loss balance—not loss alone.

Connect the ideas

How is this different from diffusion?

GAN

  • Generator creates a sample in one forward pass.
  • Discriminator provides the learning signal.
  • Training is an adversarial competition.
  • A common failure is mode collapse.

Diffusion

  • Forward diffusion gradually adds noise to real training data.
  • During training, the model learns to predict the added noise.
  • Inference begins with random noise.
  • Reverse diffusion removes predicted noise over many timesteps.
  • No discriminator is used as the primary feedback mechanism.

Reflection: Use evidence from your GAN experiment to explain how GAN training differs from forward diffusion, reverse diffusion during training, and diffusion inference.