← Entropy Atlas

Reproducible randomness

Machine-learning initialization

Seeds influence initial weights, data order, augmentation, and sampling.

Evaluation card

What “plenty” means here.

Required property
Controlled independent streams and enough repeated runs to measure sensitivity to chance.
If it fails
A single lucky seed is reported as a general improvement, or nondeterminism prevents replication.
Relevant attacker
Usually no attacker; privacy or adversarial ML systems can introduce separate requirements.
The Plenty Line
Record all relevant seeds and deterministic settings, and report distributions across multiple runs.
Why more is not automatically better
Secret high-entropy seeds do not make an experiment more valid; replicated variation does.

Where it appears

Recognize the pattern.

  • 01weight initialization
  • 02training-data shuffling
  • 03augmentation choices

The Plentropy rule

Protect the requirement, then stop.

Record all relevant seeds and deterministic settings, and report distributions across multiple runs.
Once that claim has comfortable evidence and margin, improve the next limiting factor instead of worshipping a larger entropy number.