Out-of-Distribution Generalization — Distribution Shift, Spurious Correlations, and Three Practical Remedies | MIT 6.7960 Lecture 17

🌏 中文版

Video status: Videos included. Source details

Source: based on MIT 6.7960 Fall 2024 OCW (corresponds to OCW Lec 17). Videos, slides, and assignments are all open on MIT OCW.

The gap between the i.i.d. assumption and the real world

Almost every ML textbook assumes training and test data are i.i.d. samples from the same distribution. But real deployment is never like that: the camera model changes, the hospital changes city, the corpus moves from English to legal text. A model hits 95% accuracy in-distribution, then 60% in production — not a bug, the assumption broke.

Three types of shift

The common literature splits shift into three types, each needing a different response (this taxonomy comes from the general literature, not from the structure of the Lec 17 video):

Identifying which type matters: covariate shift can be fixed by re-weighting training, label shift by calibrating predictions, concept shift usually needs re-labeling.

Spurious correlations and shortcut learning

The most common — and sneakiest — cause of OOD failure: the model latches onto a spurious correlation. The classic example: an ImageNet cow-vs-sheep classifier learns "is there grass in the background?" as a cue for cow presence; in the desert or on the beach it falls apart.

This is not a training-not-converged issue, nor an insufficient-capacity issue. ERM (empirical risk minimization) with limited capacity, given enough i.i.d. training data, picks the most predictive shortcut it can find. Avoiding it requires changing the structure of the training distribution, not a bigger model.

Three remedies

Video note: IRM, domain randomization and test-time adaptation are not in the Lec 17 transcript. The video covers adversarial examples (about 13%–36%), spurious correlations (about 49%), F1 drops under shifts such as new cameras (about 72%), worst-case / distributionally robust optimization (about 77%–85%) and diagnosing failures with generated counterfactual data (about 90%). The three remedies below are this site's supplement from the general literature.

1. Invariant Risk Minimization (IRM, Arjovsky 2019)

The idea: across multiple training environments, learn a representation Φ such that the optimal classifier w is the same in all environments — i.e. learn features that are invariant across environments. In practice, add a penalty in each environment that encourages the within-environment optimal classifier to be the same.

Practically, IRM on the DomainBed benchmark is roughly tied with plain ERM (some newer ERM variants even beat it), so it gets criticized as "conceptually elegant but limited effect." But it points to the core truth: real generalization must come from the environment structure of the training data.

2. Domain randomization / data augmentation

Since we can't enumerate all test environments, maximize environmental variation at training time: color jitter, cutout, mixup, CutMix, RandAugment, style transfer, synthetic data. The ImageNet-C / -R / -A benchmarks simulate shift via augmentation.

A cheap remedy, usually more effective than IRM. Cost: slightly longer training, and augmentation can't be too aggressive (the training distribution drifts).

3. Test-time adaptation (TTA)

At inference, use unlabeled test data to adapt the model on the fly. The classic TENT (Wang 2021) combines batch-norm stat updates with entropy minimization — a few gradient steps shift the model toward the current data distribution.

TTA needs no labels and is free at deployment, but assumes "test environment varies continuously and the model can adapt" — not a silver bullet.

Honest evaluation: the benchmark itself is biased

Recht et al. (2019) did something rigorous: re-collect the ImageNet test set with the same methodology (ImageNet-V2). SOTA models drop 11–14%. Similar conclusions from Bechtel et al. — don't take a single benchmark number as ground truth.

In practice: cross-check on at least one independent distribution (different time / geography / population) before going to production.

Minimal experiment: test-time entropy minimization

model.train()                              # enable BN stats update
for x_unlabeled, _ in test_loader:
    logits = model(x_unlabeled)
    loss = -(logits.softmax(1) * logits.log_softmax(1)).sum(1).mean()  # entropy min
    loss.backward(); opt.step(); opt.zero_grad()

Run for 1–2 epochs and the model "moves toward the test distribution's features" — covariate shift often recovers a few points this way.

Which remedy for which shift

• Small shift: more data + augmentation is enough.
• Medium shift: test-time adaptation (unlabeled) or self-training.
• Large shift: domain adaptation (a bit of target-domain data) or even retraining.
• Concept shift: almost only re-labeling — the model layer can't save you.

Course video sources

Rechecked against the live MIT OCW Fall 2024 gallery on 2026-10-10: the lecture number matches and the video is public and embeddable. Lecture numbering and topics follow OCW.

Original videos: MIT 6.7960 Fall 2024 — Lec 17. Generalization: Out-of-Distribution (OOD)

Course and recording entries:

• MIT OCW — Lec 17. Generalization: Out-of-Distribution (OOD)
• MIT 6.7960 Fall 2024 — official lecture video gallery

Checked: 2026-10-10.

Content check: verified against the video transcript (2026-10-10): read the Lec 17 transcript (Sara Beery, 1:04:41): the topic is indeed OOD generalization, but the video covers adversarial examples, spurious correlations, performance drops under shifts such as new cameras, distributionally robust optimization and counterfactual diagnosis; the transcript has no covariate/label/concept-shift taxonomy, IRM, domain randomization or test-time adaptation, so the claim that "the lecture breaks shift into three types" did not hold and was changed to "the general literature", and the three remedies now carry a note marking them as this site's supplement.

Update Log

• 2026-10-10: Added explicit video status and checked recording sources and access notes.
• 2026-10-10: Rechecked video status. Live-checked the official course page: lecture numbers and recording links match and the videos are public, so the status is now “Videos included.”
• 2026-10-10: Checked the video content against its transcript. Corrected the "lecture splits shift into three types" claim and added a note to the three remedies: they are not in the video.

References

• MIT 6.7960 OCW (Fall 2024): course home
• Arjovsky et al., Invariant Risk Minimization (2019): arXiv:1907.02893
• Gulrajani & Lopez-Paz, In Search of Lost Domain Generalization (DomainBed, 2021): arXiv:2107.00641
• Recht et al., Do ImageNet Classifiers Generalize to ImageNet? (2019): arXiv:1902.10811