AI Compass
Compass

Synthetic data

Generated rather than collected examples: when that holds, where it breaks, and why a model that only learns from models gets worse.

·2 min read·By Fachredaktion Technik
DETAIL
3 sections

The idea

Instead of collecting examples you produce them: rendered images of a component from a thousand angles, simulated sensor traces, generated sample sentences for a rare category.

When it pays

SituationSuitability
Rare defect images in manufacturingvery good, with CAD and rendering
Edge cases with known ground truthvery good
Balancing rare classesgood, with care
Replacing personal datawith limitations
Replacing expertise you do not haveunsuitable
Replacing a test setunsuitable

The routes

RoutePrincipleTypical for
AugmentationVary real dataImages, audio, text
SimulationRender a physical modelManufacturing, robotics, vehicles
Generative modelLearn from real data and generateText, images
Rule-basedGenerate from templatesForms, structured texts
Domain randomisationDeliberately unrealistic varietyTransfer to reality

The last row is counter-intuitive and well supported: when lighting, textures and backgrounds vary strongly and partly unrealistically in simulation, the model learns to ignore those factors and transfers better to real images.

  • Always mix with real data; rarely train on synthetic alone.
  • Treat the synthetic share as a hyperparameter and measure it.
  • Mark synthetic data and version it separately.
  • The test set stays real, always.

Model collapse

Variance loss across generations

σ²_n ≈ σ²₀ · (1 − 1/N)ⁿ

Repeated training on your own outputs shrinks the variance every round, and rare events disappear first.

σ²_n
variance of the distribution after n rounds of training on generated data
N
sample size per round

The practical consequence is worse than the formula suggests: the loss hits the tails of the distribution first. A model trained over several rounds on generated text first loses rare phrasings, technical terms and unusual cases, precisely what it is later needed for.

Remedy: mix real data into every round and never let its share fall below about a third.

Anonymity is not automatic

A generative model trained on real records can reproduce individual ones almost verbatim. Two checks are the minimum:

  • Nearest-neighbour distance. For every generated record, measure the distance to the most similar real one. A cluster near zero shows memorisation.
  • Membership test. Check whether the generator reveals whether a particular real record was in training.

The claim only becomes defensible with differential privacy in the generator's training:

Differential privacy

P[M(D) ∈ S] ≤ e^ε · P[M(D') ∈ S] + δ

The probability of any outcome may change by at most a factor of e to the epsilon when a single record is added or removed.

M
the generating mechanism
D, D'
two datasets differing in exactly one record
ε
the privacy budget, smaller is stricter

Only with such a guarantee and a documented ε can you argue to a supervisory authority that the generated data no longer carries a personal reference. An ε above about 10 counts as weak.

Was this page helpful?
Synthetic data