Data Augmentation, Synthetic Data, and Weak Supervision

A downstream user should be able to trace every derived example back to its origin and method. The lineage record is what makes every later check possible. Store the source ID, the generator version, the parameter set and the seed, or no one can say afterwards whether a row was observed or manufactured. The rest of this lesson is a list of what that costs: recursive collapse of the distribution's tails, memorized patient records emitted as synthetic samples, resampling that silently destroys calibration. Every one of those failures is invisible once observed and generated evidence sit in the same pile.

Augmentation creates modified views of observed examples. Synthetic data creates new records through rules, simulators, or generative models. Weak supervision creates labels from heuristics, distant signals, or multiple noisy sources. All three can expand useful training signal. All three can also repeat existing bias, invent impossible combinations, leak private examples, or teach invariances the real task does not possess. A horizontal image flip may preserve a generic object label. It can also reverse text, handedness, traffic direction, or medical laterality. A simulated failure may omit the messy precursors that appear in actual machines.

Feed generated output back into training without marking it as generated, and what happens is no longer a matter of opinion. In July 2024 Nature published the measurement. The abstract states the result plainly: “We find that indiscriminate use of model-generated content in training causes irreversible defects in the resulting models, in which tails of the original content distribution disappear.” The six authors named the effect model collapse. Then they measured it. They fine-tuned OPT-125m recursively on its own wikitext2 output and tracked perplexity. With real data, “The original model fine-tuned with real wikitext2 data obtains 34 mean perplexity, from the zero-shot baseline of 115”. With no original training data retained, “training with generated data allows us to adapt to the underlying task, losing some performance, from 20 to 28 perplexity points”.

The finding is not confined to text, or to one laboratory. An independent group at Rice University found the same thing in state-of-the-art generative image models, in a 2024 paper called Self-Consuming Generative Models Go MAD. They named the condition Model Autophagy Disorder and reported that “appreciable MADness arises in just a few generations”. Two research communities, two media, one mechanism: the rare cases go first.

The pipeline should therefore record source example, generator, parameters, label rule, and version. Synthetic evidence must stay distinguishable from independently observed evidence during evaluation and incident analysis. As the collapse work shows, it has to stay distinguishable during the next training run too.

Whether the label survives is task-specific, not visually obvious. Three Google researchers built an entire dataset out of this one failure. PAWS — Paraphrase Adversaries from Word Scrambling, 2019 — contains 108,463 human-labeled paraphrase and non-paraphrase pairs with high lexical overlap, produced by controlled word swapping and back translation. The canonical pair is “flights from New York to Florida” against “flights from Florida to New York”. Every token is preserved. The meaning is reversed. Systems that had learned paraphrase from ordinary corpora fell apart on it: “State-of-the-art models trained on existing datasets have dismal performance on PAWS (<40% accuracy); however, including PAWS training data for these models improves their accuracy to 85% while maintaining performance on existing tasks.” Google's own release of the data gives the composition. PAWS-Wiki holds 49,401 labeled training pairs, 44.2% of them paraphrases, alongside 656k noisily labeled pairs. PAWS-QQP holds 11,988 training pairs, 31.3% paraphrases. A paraphrase generator used as an augmenter runs that same transformation — without the labels that would tell you when it went wrong.

Methods differ in whether they change inputs, create new examples, or infer labels. The last of the five, resampling and interpolation, has the longest published record and the sharpest warning attached to it. SMOTE dates from 2002. Its own description: a “method of over-sampling the minority class involves creating synthetic minority class examples”. Twenty years later, a study in JAMIA measured what that does to a clinical risk model. Monte Carlo simulation, plus an ovarian-cancer case study, and the result was blunt: “The use of random undersampling, random oversampling, or SMOTE yielded poorly calibrated models: the probability to belong to the minority class was strongly overestimated.” There was no compensating gain in AUC. Their conclusion is the sentence to remember before reaching for the technique: “Outcome imbalance is not a problem in itself, imbalance correction may even worsen model performance.” A later simulation extended the same finding to machine-learning algorithms and to MIMIC-III. Interpolating between observed feature vectors changes the optimization problem. It does not add independent real-world coverage. And it can cost the calibration the model was supposed to deliver.

Begin from the failure or coverage gap the method is supposed to address. Each step below exists because a published result showed what its absence costs. Provenance, because model output fed back unmarked erodes the tails of the distribution. Semantic review, because word-order swaps that preserve every token reverse the label. Independent real-data validation, because risk models rebalanced with SMOTE keep their discrimination and lose their calibration. Approve a generator the way a measurement instrument is approved — against observations it did not produce.

Multiple labeling rules may make the same mistake because they share a source, keyword list, or upstream model. Majority vote then overstates independent evidence.

Measure rule coverage, overlap, conflict, class tendency, and performance on a manually reviewed sample. Preserve each rule's output rather than only the combined label. Rules can be useful as features, candidate generators, or training targets under different protocols, and a rule that performs poorly alone may still contribute complementary information. Version the rules and recompute weak labels when logic changes. Mix rule generations without provenance and you cannot tell whether things got better or worse.

Run that discipline properly and the payoff is measurable in production, not only on benchmarks. In 2019 Google deployed programmatic labelling functions on three production classification tasks, in a system called Snorkel DryBell. The paper reports: “On three classification tasks at Google, we find that Snorkel DryBell creates classifiers of comparable quality to ones trained with tens of thousands of hand-labeled examples, converts non-servable organizational resources to servable models for an average 52% performance improvement, and executes over millions of data points in tens of minutes.” Google's own account of the deployment puts the equivalence in counts: “we achieved the same predictive accuracy training a model with Snorkel DryBell's labels as we did when training that same model with 12,000 and 80,000 hand-labeled training data points”, and reports that cross-feature transfer “boosted our performance by an average 52% on the benchmark datasets we created”. Those numbers are the reward for treating the rules as an instrument to be characterized, not as an oracle to be averaged.