fall_2021

Super-resolution is a hot topic in current day Machine Learning.  The origin of the methodology dates back to applications in seismic imaging. I discuss the evolution from the early days and highlight some papers which have given new theoretical insights along the way. I illustrate the bridge between traditional convex optimization and current day convolutional neural nets. Along the way I show some examples where we have used this for current day applications in seismic imaging.

For many real-life phenomena one may assume that the units of observation, typically patients, transition through a set of discrete states on their way towards an absorbing state. The states often constitute various stages of a disease, from perfect health through various stages of dementia for example. Multi-state models are a class of statistical models which allow us to study the time spent in different states, the probability of transitioning between states, and the relationship between these quantities and covariates of interest. In many applications the transition times between states are not observed exactly; instead, the current state of the patients is queried at arbitrary times. The transition times are therefore interval censored, and this makes inference and modelling challenging. Most current approaches are based on the Markov assumption, for example the simplest parametric model available - the time-homogeneous Markov model. Here, we propose a new, general framework for parametric inference with interval censored multi-state data. Our models allow non-Markovian behaviour. I will present the framework and an algorithm for the automatic construction of the likelihood function, along with real-data examples. This talk is based on joint work with Marthe Aastveit and Nils Lid Hjort.

Cards are drawn, one at a time, with replacement, from a deck of n cards. I study the total time W_n needed until we have seen all n cards, via different perspectives, along with a Gumbel limiting distribution. Various non-trivial identities, involving different perspectives for moments and Laplace transformations, are found as corollaries. These findings are also used to estimate the number of different cards,if uknown. If I needed to sample 133 words from a document, before I had 50 different words, what is the vocabulary size for the document? How many words did Shakespeare know (including those he never used in his writing)? 

An Abels Tårn podcast about some of these themes, which attracted a fair amount of inspired comments and guesses from the public (specifically, finding the mean of W_n above, for the case of n = 52 cards), can be found on the Abels Tårn website, July 2021, as a conversation with Torkild Jemterud, Jo Røislien, and myself. 

The talk is elementary and discusses empirical modelling of single variables with insurance losses as example. There are in such cases little or no theory to go on, and the amount of data is many situation quite scarce. Why do we so often limit ourselves to fit two-parameter families? It will be suggested that it may be a good idea to work with more flexible models with three or four parameters and that this may provide a nice framework for automating the entire procedure for the computer to work alone. Sure, with little data the parameters may be unstably estimated, but that may not apply equally to the distributions they define. Many-parameter families suitable for insurance losses will be reviewed with some simple asymptotics in an example allowing this and with Monte Carlo to throw light on the issue in other cases.

In this talk I will discuss the variational form of Bayes theorem by Zellner (1988). This result is the rationale behind the variational (approximate) inference scheme, although it is not always that clear in modern presentations. I will discuss two applications of this results. First, I will show how to do a low-rank mean correction within the INLA framework (with amazing results), which is essential for the next generation of the R-INLA software currently in development. In the second one, I will introduce the Bayesian   learning rule, which unify many machine-learning algorithms from fields such as optimization, deep learning, and graphical models. This includes classical algorithms such as ridge regression, Newton's method, and Kalman filter, as well as modern deep-learning algorithms such as stochastic-gradient descent, RMSprop, and Dropout.

The first part of the talk is based on our recent research at KAUST, while the second part is based upon \texttt{arxiv.org/abs/2107.04562} with Dr. Mohammad Emtiyaz Khan, RIKEN Center for AI Project, Tokyo.

Marginal maximum likelihood estimation of longitudinal latent variable models for ordinal observed variables is challenging due to the high latent dimensionality required to accurately model residual dependencies for repeated measurements. We use second-order Laplace approximations to the high-dimensional integrals in the marginal likelihood function for longitudinal item response theory models and implement an efficient estimation method based on the approximations. The method is illustrated with items from the Montreal Cognitive Assessment, administered at four time points in a Hong Kong study of aging and well-being. We discuss the limitations of the proposed estimation method and outline a potential extension to the approach that uses a dimension-reduction technique.