Corinna Cortes

Life and career

Corinna Cortes came to machine learning from physics, which is not unusual, and then stayed to run one of industry's most productive research labs, which is. Her name is attached to a single 1995 paper that changed what classifiers looked like for a decade, but the shape of her career — Copenhagen, Rochester, Bell Labs, Google — says something more interesting about how the field actually developed than the one-line summary suggests.

She trained in physics in Denmark, at the University of Copenhagen, before crossing to computer science for her doctorate at the University of Rochester. The physics training left visible traces: a comfort with empirical scaling arguments, an instinct for asking how a quantity behaves as you push a parameter, and a tolerance for messy measurement that a purely theoretical education does not always instill.

The formative period was Bell Labs, and later AT&T Labs, in the 1990s. The Holmdel and Murray Hill groups of that era held an implausible concentration of people who would define statistical learning: Vladimir Vapnik, Yann LeCun, Yoav Freund, Robert Schapire, Isabelle Guyon, and others, all working on the same intractable industrial problems. The central benchmark was handwritten character recognition — reading zip codes and cheque amounts — and Cortes worked directly on it. This matters for understanding the SVM: it was not conceived as an abstraction. It was built in a building where people had a hard, noisy, sixty-thousand-example classification problem on their hands and a competing method (convolutional networks, two doors down) setting the bar.

Cortes also worked on a very different problem at AT&T: the extraction of statistical signatures from enormous transactional data streams — telephone call records, at a scale where you could not hold the data in memory and could not afford more than a pass or two over it. That work, including the Hancock system for computing per-customer signatures over massive streams, was used for fraud detection and marketing analytics. It is the least-cited part of her record and arguably the most predictive of her later career, because it was industrial machine learning at web scale a decade before the web had that scale.

She joined Google, where she has led Google Research New York for many years, growing it from a small group into one of the company's principal research centers. Running a lab of that size is a different job from doing research, but she kept publishing throughout, largely in a long-running collaboration with Mehryar Mohri on learning theory with immediate practical teeth: learning kernels, domain adaptation, importance weighting under sample selection bias, classification with a rejection option, and structural approaches to learning neural network architectures.

Cortes and Vapnik received the ACM Paris Kanellakis Theory and Practice Award for the support vector machine — the same award Freund and Schapire hold for boosting, and for the same reason: a theorem that turned into a tool everyone used. She is also a serious competitive distance runner, which anyone who has worked with her tends to mention within about five minutes.

Key contributions

**The soft-margin SVM.** The hard-margin support vector machine — find the separating hyperplane that maximizes the distance to the nearest point — is elegant and nearly useless, because it requires that the classes actually be separable in the feature space you chose. One mislabeled example, or one genuinely ambiguous point, and the optimization problem is infeasible. In practice it also means that the solution is exquisitely sensitive to the worst point in the dataset, which is a terrible property.

Cortes and Vapnik's 1995 "Support-Vector Networks" introduced **slack variables**. Each training point gets a nonnegative slack ξᵢ measuring how far it is allowed to violate the margin, and the objective becomes: minimize ‖w‖²/2 + C·Σξᵢ subject to the relaxed margin constraints. The parameter C sets the exchange rate between a wide margin and a clean fit. What makes this more than a hack is that the resulting problem is still a convex quadratic program, its dual is still expressed purely through inner products (so kernels still work), and the dual now has *box constraints* — the multipliers are capped at C — which has a clean interpretation: no single point, however badly placed, can dominate the solution. Points at the cap are the ones on the wrong side; the geometry stays interpretable.

The consequences were immediate and large. Soft margins are what made SVMs deployable on real data, and combined with the kernel trick they gave the field a classifier with a convex objective, a small number of tunable knobs, a capacity-control story rooted in VC theory rather than folklore, and excellent out-of-the-box accuracy. For roughly the decade from the late 1990s, the SVM was the default answer to "I have a classification problem."

**Large-scale stream analytics.** The signature-extraction work at AT&T addressed a problem statisticians of the era largely ignored: computing and maintaining a per-entity statistical profile over a stream too large to store, updating it incrementally, and doing so for tens of millions of entities. The engineering discipline this required — one pass, bounded memory, approximate but stable summaries — is now standard, and it was being done for fraud detection on telephone networks in the 1990s.

**Learning theory for practical failure modes.** With Mohri, Cortes has produced a long series of papers on the ways deployed learning goes wrong. *Domain adaptation*: when the training and test distributions differ, what discrepancy measure between them actually controls the excess risk, and what reweighting is justified? *Sample selection bias and importance weighting*: importance weights have unbounded variance in general, so what guarantees survive? *Learning kernels*: instead of picking a kernel by cross-validation, learn a combination of base kernels from data, with Rademacher-complexity bounds showing what that extra flexibility costs. *Learning with rejection*: give the classifier the option to abstain, and jointly learn the classifier and the rejection rule with a calibrated surrogate loss. *AdaNet*: adaptively grow a neural architecture, adding units and layers only when a data-dependent complexity bound says the addition pays for itself.

That last cluster is where her two identities meet — a lab director who sees what breaks in production, and a theorist who insists on a bound before shipping a fix.

In battle

Cortes's profile is that of an extreme specialist with an unusually deep trough: mean 24.0, median 20, fifty-four problems at or below 20, and exactly one problem above 80. This is a figure you must aim carefully.

`classification` at 65.3 is her only strong category, and it is very strong. Her top problem, **P025 — Sixty thousand digits** at 88, is nearly autobiographical: this is the Bell Labs handwritten-digit benchmark, at the scale she worked it, with the method she built for it. The game's explanation is admirably honest about why she is not at 98 — the punchline of that benchmark is that fixed-kernel SVMs eventually ceded the last fraction of a percent to networks that learned their own features. She was in the room, on the data, with a contending method, and the method came second.

**P258 — The inbox under siege** (74) is text classification with a linear margin classifier, which is close to what spam filtering actually looked like in the early 2000s. **P026 — A committee of weak learners** (68) puts her in the room for boosting without owning it — a good, defensible score that will lose to Schapire, Freund, or Bartlett. **P070 — Find every face** (60), **P020 — The smooth curve, honestly** (58), **P018 — p = 20,000, n = 200** (54), **P022 — Predict, or explain?** (48), and **P278 — All your neighbors are equidistant** (48) fill out a band of decent-but-beatable scores across supervised learning generally.

The weaknesses are stark and instructive. **P143 — The coefficient that flips sign** at 3 is the lowest score in her set and one of the lowest anywhere: margin maximization has no machinery for confounding, mediation, or collider bias, and the game refuses to award credit for general statistical sophistication. **P118 — The factor you can't keep changing** (8) is experimental design — her `experimental-design` category sits at 10.0, a genuine hole. **P142 — Predict the ore grade underground** (7) is geostatistics and kriging; **P173 — Can these codeword lengths even exist?** (7) is coding theory; **P244 — The compiler that beats the coder** (8) and **P245 — A computer shared by fifty** (6) are systems, where she sits at 7.0.

There is a subtler pattern worth naming. `perception` at 22.4 across nineteen problems and `nlp` at 22.1 across eighteen look low for someone who ran a major industrial AI lab, and they are low for exactly the reason her battle identity states: kernel methods do not scale to web-sized data, and they do not learn representations. On perceptual and language problems, the era that succeeded her method dominates. Her long-run industrial contributions — stream analytics, domain adaptation, architecture search — are diffuse and shared, and the scoring rewards the sharp, attributable result.

Play Cortes on a supervised classification problem where the classes overlap, the data is moderately sized, and someone needs a defensible decision boundary today. Anywhere else, she is outgunned.