Research Group Leader
For more information about our work, see also our group's homepage.
See also my CV for more information on myself.
My work concerns
Probabilistic Uncertainty over the Result of Computations
Intelligence is the ability to act under uncertainty. It exists on a broad range, not just of physical, but also of computational scales: From simplistic ideas like gradient descent, which may be a microbe's strategy to get closer to a source of nutrients, to an adult human's reasoning about career goals. Much of modern research in machine learning and artificial intelligence aims for the top of this hierarchy: algorithms capable of building highly structured models, and taking complicated decisions, at high computational cost. I believe that there is still plenty of room for improvement left at the bottom, too.
Algorithms for the bottom end of the intelligence hierarchy are those constructed by numerical mathematics. They are methods that estimate the result of intractable computations: Optimizers estimate the location of (local or global) extrema. Quadrature methods estimate values of integrals. Differential equation solvers estimate curves fulfilling constraints on their derivatives. In this sense, all these methods can be seen as inference algorithms, estimating latent quantities from the result of tractable computations.
These algorithms are the building blocks for the more complex top level intelligence. So they have to be modular, to be re-usable. They have to be robust, because their failure may cause big problems upstream. And of course they have to be cheap. In my work, I try to address theses requirements. Our analysis starts with the observation that many existing, basic numerical methods can be interpreted precisely as maximum a-posterior estimators under Gaussian priors and likelihoods. This includes real classics, like the conjugate gradient algorithm [Hennig, 2014], the BFGS rule [Hennig & Kiefel, 2013], Gaussian quadrature (including Romberg's rule and others) and the family of Runge-Kutta methods [Schober, Duvenaud & Hennig, 2014]. Taking this interpretation as a starting point, we extend the established methods to address contemporary challenges of large-scale learning: Modelling noise in observations and building structured priors for improved performance on specific tasks. A particularly promising area of generalization is the sharing of information among related problems, and the propagation of information from one computational step to the next. Over time, we plan to extend this notion into a general purpose framework for uncertain computation.
Here is an older selection of some of my work. See "publications" for pdfs and detailed citations, my CV (link above) for more recent information, and http://probabilistic-numerics.org for a frequently updated overview of our emerging community:
Gauss-Markov Runge-Kutta methods for stochastic models of ODE solutions
Runge-Kutta methods, the bedrock of numerical ODE solution methods, can be interpreted as the posterior mean of a specific kind of linear state-space model under a Wiener drift [Schober, Duvenaud & Hennig. Probabilistic ODE Solvers with Runge-Kutta Means. NIPS 2014]. This insight can be used to construct methods assigning probability measures over the solution of an ODE. This measure can be visualised as an uncertainty [Schober et al., Probabilistic Shortest Path Tractography in DTI using Gaussian Process ODE solvers, MICCAI 2014], and used to control the computational cost of ODE solvers [Hennig & Hauberg, Probabilistic Solutions to Differential Equations and their Application to Riemannian Statistics, AISTATS 2014].
propagation of uncertainty for radiation treatment planning
together with Mark Bangert at the German Cancer Research Centre, we are developing a toolchain for the analytic propagation of setup uncertainties through the optimization chain used in radiation therapy treatment planning, aiming to find more robust treatment plans that reduce the incidence of complications and side-effects in radiation tumor therapy. See Bangert, Hennig & Oelfke "Analytical probabilistic modeling for radiation therapy treatment planning Physics in Medicine and Biology" 58(16) 5401-5419.
quadratic optimization under noise
Stochastic gradient descent is still the dominant algorithm for the training of many online learning algorithms, like neural networks. All just because more elaborate ideas, like quasi-Newton methods, cannot deal with noise? See what can be done about that: Hennig. "Fast Probabilistic Optimization from Noisy Gradients". ICML 2013
nonparametric quasi-Newton methods
Did you know that BFGS is a least-squares regressor? See what happens when you make it nonparametric: Hennig & Kiefel. "Quasi-Newton methods, a new direction". ICML 2012
information-efficient experimental design
When optimizing experimental parameters in search of a global optimum, algorithms shouldn't try evaluating close to the optimum. They should try to evaluate where they expect to learn most about the optimum. Hennig & Schuler, "Entropy Search for Information Efficient Global Optimization". JMLR 13 (2012).
optimal Bayesian reinforcement learning
Probability theory offers a uniquely coherent view on the infamous exploration/exploitation tradeoff: From the Bayesian view, reinforcement learning is about modelling the effect of possible future observations on the optimality of decisions taken in the present. In general, this decision process is intractable. But under Gaussian process assumptions (which, depending how on look on it, is either a quite general, or a quite limited set of assumptions), the right answer moves within reach of numerical analysis. Hennig, "Optimal Reinforcement Learning for Gaussian Systems", NIPS 2011
kernel topic models: fast inference in dependent Dirichlet models
Topic modelling is a very popular area of machine learning at the moment. Documents come with metadata, and topics change over time, and from document to document depending on the author, the subject, and many other features. The probabilistic extension of topic models that allows modelling such effects requires an algorithmic link between discrete distributions and continuous domains, often realised as a set of "dependent Dirichlets". We pointed out how to do this, in a numerically extremely efficient way. Hennig, Stern, Herbrich and Graepel, "Kernel Topic Models". AISTATS 2011.
Bayesian tree search
Tree search, finding the optimal leaf of a tree, is exponentially hard in the depth of the tree, because trees are exponentially big in their depth. But what happens during that exponentially long search? If you have a probabilistic belief over the value and location of the optimal leaf, and get one more observation of one individual leaf's values? Shouldn't updating the belief cost only linear time? It does. Hennig, Stern and Grapel. "Coherent Inference on Optimal Play in Game Trees". AISTATS 2010