Who is Daphne Koller?
Daphne Koller is a computer scientist and entrepreneur working at the intersection of artificial intelligence, biology and education. She is the founder and CEO of insitro and a cofounder of Coursera and Engageli. Her academic specialty is building models that reason from incomplete information and learn relationships between variables. [1]
She was born in Jerusalem on 27 August 1968. [21]
Education and early research
She earned her undergraduate degree at the Hebrew University of Jerusalem in 1985 and her master’s in 1986. She completed a Stanford computer science doctorate in 1993 and undertook postdoctoral research at the University of California, Berkeley, from 1993 to 1995. The MacArthur Foundation later recognized her work with a fellowship in 2004. [6]
Her dissertation, From Knowledge to Belief, was supervised by Joseph Halpern. Her early research examined representing uncertain knowledge and making decisions using probability, logic and game theory. The problem was not simply storing facts: reasoning also had to work when a description of the world was insufficient to establish certainty. [7]
Stanford, research and student development
She joined Stanford’s computer science department in September 1995. She held the Rajeev Motwani professorship and a courtesy appointment in pathology, reflecting her research’s growing connection with medicine. This period brought together work on knowledge representation, learning and applications to complex data. [2]
In 2001 she initiated Stanford’s summer research programme for computer science undergraduates, with faculty mentorship. She received the Cox Medal in 2003 for fostering undergraduate research. The programme offered entry into scientific projects before doctoral study and complemented her classroom teaching. [15]
Bayesian networks and structured learning
With Avi Pfeffer, she presented object-oriented Bayesian networks at UAI 1997. The method describes complex systems through related objects, reusable classes and hierarchies. It avoids rebuilding a model from scratch for each similar component and exploits that structure to organize probabilistic inference. [8]
At UAI 2002, Ben Taskar, Pieter Abbeel and Koller developed discriminative models for relational data. Their approach jointly classified connected entities, such as linked web pages, rather than ignoring dependencies. Markov networks provided a representation without the acyclicity constraint imposed by directed Bayesian networks. [9]
The 2003 paper Max-Margin Markov Networks, with Taskar and Carlos Guestrin, combined margin-based learning with dependencies between labels. It studied handwriting recognition and hypertext classification. Its methodological contribution was to exploit output structure when deciding each label independently would waste information about the whole. [10]
Computational biology and medicine
In 2003 she worked with Eran Segal, Nir Friedman and other researchers on Module Networks. The method analysed gene-expression data to group genes and propose which regulators acted under particular conditions. Its results took the form of testable hypotheses, connecting statistical learning with experimental questions about cellular organization. [11]
A 2004 paper with Segal, Friedman and Aviv Regev extended this approach to cancer. It integrated 1,975 microarrays spanning 22 tumour types to study activated and deactivated sets of genes. This enabled comparisons across diseases and tissues without reducing each tumour to an independent gene list. [12]
She was also a senior author of PhysiScore, research published in 2010 with Anna Penn and led by Suchi Saria. The study of 138 premature infants combined early-life signals to estimate complications. Stanford expressly stated that larger-population validation and assessment of its effect on clinical decisions were required before commercial use. [13]
The textbook Probabilistic Graphical Models
With Nir Friedman, she published Probabilistic Graphical Models: Principles and Techniques in 2009. The book connects representation, inference and learning in Bayesian networks, Markov networks and dynamic or relational models. It provides a framework for constructing interpretable models and calculating consequences of available information, with different reading pathways depending on the student’s background. [14]
Koller brought some of this teaching into Stanford Online’s Coursera specialization. The introductory course distinguishes directed and undirected networks and covers influence diagrams and decisions. Programming assignments make constructing and using representations part of the learning objective, alongside understanding their definitions. [17]
Coursera and Engageli
Andrew Ng and Koller launched Coursera in April 2012. The initiative brought university courses to an international audience through the internet. It grew from their teaching experience and the possibility of distributing materials and activities at a scale different from the traditional physical classroom. [16]
In 2020 she cofounded Engageli with Dan Avida, Serge Plotkin and Jamie Nacht Farrell. The platform launched to support participatory digital classes during the expansion of remote education. It combined group work and interaction tools for teachers and students, a different emphasis from individual viewing of recorded videos. [18]
From Calico to insitro
From 2016 to 2018 she was Calico’s first chief computing officer. At the Alphabet health-research company, she worked on applying machine learning to biological datasets. This step brought her computer science research experience into organizations whose work requires connecting computation, experimentation and biomedical development. [3]
Explaining insitro’s origins, Koller emphasized that early applications of AI to biology had encountered scarce or unsuitable data. Improved measurements and experimental tools enabled another strategy: generating and analysing useful data to investigate disease causes. She also highlighted the difficulty of getting specialists from different disciplines to work with a shared language. [4]
In 2019 insitro announced a collaboration with Gilead to investigate the liver disease then called NASH. The agreement envisaged identifying potential targets through models and data, with Gilead able to advance up to five. The pharmaceutical company would handle the corresponding chemistry and development; identifying a target was not equivalent to having a treatment. [19]
In October 2026 insitro announced an agreement with AstraZeneca covering MASH and cardiomyopathies. Insitro would investigate targets and candidates; AstraZeneca would handle clinical development, regulation and commercialization. The announcement placed its first liver programmes’ clinical entry in early 2027: a development forecast, rather than evidence of efficacy in patients. [20]
Recognition and influence
The ACM Prize in Computing for 2007 recognized her combination of relational logic and probability, applicable to robotics, economics and biology. Its significance lies in providing general tools for different problems while retaining an explicit representation of relationships and uncertainty. [21]
The 2019 ACM–AAAI Allen Newell Award highlighted her contributions to probabilistic models, their application to biology and health, and wider educational opportunity. This combination explains the breadth of her career: mathematical methods, teaching and the organizations she built are not entirely separate stages. [22]
In 2023 she was elected to the United States National Academy of Sciences. The recognition added an institutional dimension to a career spanning computing and the life sciences. [23]
Reliable perception for robots operating around people
Koller coauthored research with Anna Petrovskaia, Sebastian Thrun and Oussama Khatib on reliable robot-position estimation. The GRAB method incorporates error bounds and is studied in indoor navigation and tactile manipulation, where incorrect perception can lead to unsafe decisions. [5]
This is a specific prevention connection: before deciding how to move, a machine needs to know its position and how wrong that estimate could be. An apparently accurate robot does not provide the same information as a system that characterizes its estimation limits.
From perception tests to workplace safety
The work establishes research into perception reliability; it does not demonstrate preventive deployment in factories or make every robot using probabilistic inference safe. Koller’s contribution concerns this particular mathematical and experimental problem.
Assessing a workplace application would require checking sensors, occlusions and conditions that invalidate the model, alongside movement safeguards and human interaction. A correctly estimated position is one component of system safety, rather than a complete safety assessment.
