# Dropout reference

Run `python examples/dropout.py` or append `--json` for the exact eight-mask teaching payload. Python 3.10+; no external packages.

Source: Srivastava et al., JMLR 15(56), 2014, Sections 4–5 and Figure 2. Regions `mask`, `scaling`, and `ensemble` are canonical source excerpts for the central synchronizer. `keep` means retention probability, unlike APIs that accept a drop probability.

The original training convention uses r*h with test-time scaling p. The optional inverted convention uses r*h/p with identity inference. Original p=0 is defined; inverted p=0 is rejected. All masks are explicit or exhaustively enumerated, so outputs are deterministic. The sigmoid enumeration illustrates arithmetic nonlinear expectation, not a reproduction of the paper's empirical model-average comparison.

This is not a trained network, classification experiment, checkpoint, random accuracy estimate or latency benchmark. The learned hidden features, minibatch optimizer, max-norm constraints and datasets are omitted. Unit tests verify probability mass, expectation/variance identities, boundaries, fixed-mask finite differences and JS/Python parity: `node --test test/dropout.test.mjs`. Source crops are hash-pinned in `public/images/papers/dropout/source-crops.json`.
