Daphne Koller

Life and career

Daphne Koller has had three careers, and the striking thing is that the same idea runs through all of them: if you want to reason about a complicated world, first write down honestly what depends on what.

She was born in Jerusalem in 1968 and was extraordinarily accelerated as a student, entering the Hebrew University at seventeen and finishing a master's degree by eighteen. She came to Stanford for a PhD, completing it in 1993 under Joseph Halpern in a tradition concerned with the logic and semantics of reasoning under uncertainty — a background in formal knowledge representation rather than in statistics or engineering. A postdoc at Berkeley with Stuart Russell followed, and in 1995 she joined the Stanford faculty, where she spent two decades and became one of the central figures of the probabilistic-graphical-model era of machine learning. Along the way she received a MacArthur Fellowship and the ACM Prize in Computing, and was elected to the national academies.

Her Stanford period defined the field's vocabulary. Judea Pearl had supplied Bayesian networks as a representation; Koller's program was to make them into an engineering discipline — how to represent structured and relational domains, how to do inference when the graph is too dense for exact methods, how to learn both the parameters and the structure from data, and how to combine probabilistic structure with discriminative training. The 2009 textbook she wrote with Nir Friedman, *Probabilistic Graphical Models: Principles and Techniques*, is a thousand-page manual for that discipline and remains the reference.

The second career began in 2011, when she and Andrew Ng put Stanford courses online and found tens of thousands of people taking them. They founded Coursera in 2012, and Koller served as co-CEO and then president. It is worth noting how directly this grew from her teaching: her graphical models course was among the first offerings, and the motivation was that a body of technique this useful should not be gated by admission to one university.

The third career is biology. Koller had worked on computational biology throughout her Stanford years — most notably the module-network approach with Eran Segal and Nir Friedman, which inferred regulatory modules and their controlling programs from gene expression data, a graphical-model formulation applied to a genuine biological question and published in the genetics literature. In 2016 she became chief computing officer at Calico, Alphabet's longevity research company, and in 2018 she founded insitro, where she is CEO. Insitro's thesis is a machine-learning thesis about drug discovery: the bottleneck is not compound screening but the absence of large, consistent, well-controlled biological datasets, so build the wet lab to *generate* data at the scale and quality that learning requires, rather than mining whatever happens to exist. That is a bet on experimental design as the limiting resource, which is a very Koller position.

Key contributions

**Probabilistic graphical models as an engineering discipline.** The central object is a factorization: a joint distribution over many variables represented compactly by a graph, either directed (Bayesian network, factorizing into conditional distributions given parents) or undirected (Markov random field, factorizing into potentials over cliques). The payoff is that a joint over hundreds of binary variables, which would require an astronomically large table, is specified by a modest number of local factors. Koller's textbook and research systematized the three hard questions this raises. *Representation*: what independence assumptions does a given graph encode, and how do you build graphs for domains with repeated structure. *Inference*: exact inference by the junction-tree algorithm costs exponential time in the treewidth of the graph, which is fine for chains and trees and hopeless for realistic dense models — so when treewidth explodes you need approximate methods, loopy belief propagation, variational approximations, or MCMC, and you need to understand what each one is actually approximating. *Learning*: estimating parameters from data, including with missing values via EM, and the much harder problem of learning the graph structure itself, where the search space is superexponential and the scoring function must penalize complexity.

**Probabilistic relational models.** Standard Bayesian networks assume a fixed set of variables. Real domains have objects, classes, and relations — patients, doctors, hospitals; students, courses, grades — with structure that repeats across instances. With Avi Pfeffer and others, Koller extended graphical models to relational settings, so that a template defined at the class level unrolls into a ground network over whatever objects the database actually contains, with parameters shared across instances. This is the foundation of statistical relational learning, and the same "define a template, tie parameters across instantiations" logic recurs throughout modern ML.

**Max-margin Markov networks (2003, with Taskar and Guestrin).** Generative graphical models are trained to fit the joint distribution; SVMs are trained to maximize a margin but assume independent outputs. M³Ns combine them: keep the graphical model's factorization over structured outputs, but train with a large-margin objective over the exponentially many possible output configurations, using the factorization to represent the exponentially many constraints compactly. This is one of the founding results of structured prediction.

**Active learning (with Simon Tong, 2001).** For support vector machines, the most informative unlabeled point to query is the one that most reduces the version space — approximately, the point closest to the current decision boundary. Simple, principled, and one of the most-used active-learning heuristics.

**Computational biology.** Module networks and related work, in which the graphical model is used not merely to predict but to represent a hypothesis about biological mechanism — which regulators control which sets of genes, under which conditions.

In battle

Koller's sheet is unusually well-shaped: mean 33.1, median 30, and only twenty-one problems at or below 20. She is not a spike; she is a broad probabilistic reasoner who is strong wherever structure and uncertainty meet.

Her peak, "The web of symptoms" at 92, is the medical-diagnosis network — encoding a large joint over diseases and symptoms compactly, choosing between exact junction-tree inference and approximations when the treewidth blows up, and learning conditional tables from sparse patient data. The game's explanation notes she is the *second-generation* heir: the formalism is Pearl's, and in the year the historical problem is set she was a teenager. Given the formalism, she executes the build as well as anyone alive, and the biomedical framing is native terrain rather than analogy.

Her two strongest categories say something important that students will miss from the summary line. Causality at 74.3 and reinforcement learning at 74.0 are both far above her overall mean. "Which one is the cause?" (68) and "Act on what you cannot see" (68) — the latter a partially observed decision problem — reward the fact that a graphical model with interventions is a causal model, and that influence diagrams and POMDPs are the same representation extended with decisions and utilities. "The policy learned from old charts" at 80 is offline policy learning from observational medical records, which sits precisely at the intersection of her formal training, her biomedical work, and her insistence that confounded logged data must be handled explicitly. Very few figures on this roster can play causal problems at all; Koller is one of the strongest.

Her other wins follow the same logic: "Which examples deserve labels?" (84) is her active learning result; "Where to place the sensors" (76) is experimental design as an information-maximization problem over a structured model; "The tilted scan" (79) reflects structured models over images, where deformations are handled by an explicit model of spatial dependency rather than by data augmentation; "The posterior at web scale" (63) is approximate inference under a computational budget.

The losses are clean and are mostly *not* statistical. Her floors are systems (7.5), information (9.0), and computability (10.5). "Small tools, loosely joined" (6) and "A computer shared by fifty" (9) are operating-system design; the packet's language is blunt about there being no bridge from graphical-model inference to shell pipelines or scheduler design. "Compress without knowing the source" (8) and "The optimal codebook" (10) are source coding and quantization — information theory proper, which her framework uses as a scoring criterion but has never contributed to. "The equation with no algorithm" (9) is undecidability. And "Beat the world champion" (10) is worth dwelling on: Deep Blue was exhaustive search with a hand-tuned evaluation, containing no uncertainty, no latent variables, and nothing to infer — the single environment in which her entire toolkit has no purchase.

Play Koller when the world has many interacting variables, some of them unobserved, and you need to reason about what would happen if you intervened. Bench her when the problem is a machine, a code, or a proof.