Working paper
Robust Variational Inference
The classical parametric and semiparametric Bernstein – von Mises (BvM) results are reconsidered in a non-classical setup allowing finite samples and model misspecification. In the case of a finite dimensional nuisance parameter we obtain an upper bound on the error of Gaussian approximation of the posterior distribution for the target parameter which is explicit in the dimension of the nuisance and target parameters. This helps to identify the so called critical dimension p of the full parameter for which the BvM result is applicable. In the important i.i.d. case, we show that the condition “p^3 / n is small” is sufficient for the BvM result to be valid under general assumptions on the model. We also provide an example of a model with the phase transition effect: the statement of the BvM theorem fails when the dimension pn approaches n^{1/3} . The results are extended to the case of infinite dimensional parameters with the nuisance parameter from a Sobolev class.
Recently proposed distance dependent Chinese Restaurant Process (ddCRP) generalizes extensively used Chinese Restaurant Process (CRP) by accounting for dependencies between data points. Its posterior is intractable and so far only MCMC methods were used for inference. Because of very different nature of ddCRP no prior developments in variational methods for Bayesian nonparametrics are appliable. In this paper we propose novel variational inference for important sequential case of ddCRP (seqddCRP) by revealing its connection with Laplacian of random graph constructed by the process. We develop efficient algorithm for optimizing variational lower bound and demonstrate its efficiency comparing to Gibbs sampler. We also apply our variational approximation to CRP-equivalent seqddCRP-mixture model, where it could be considered as alternative to one based on truncated stick-breaking representation. This allowed us to achieve significantly better variational lower bound than variational approximation based on truncated stick breaking for Dirichlet process.
Recently proposed distance dependent Chinese Restaurant Process (ddCRP) generalizes extensively used Chinese Restaurant Process (CRP) by accounting for dependencies between data points. Its posterior is intractable and so far only MCMC methods were used for inference. Because of very different nature of ddCRP no prior developments in variational methods for Bayesian nonparametrics are appliable. In this paper we propose novel variational inference for important sequential case of ddCRP (seqddCRP) by revealing its connection with Laplacian of random graph constructed by the process. We develop efficient algorithm for optimizing variational lower bound and demonstrate its efficiency comparing to Gibbs sampler. We also apply our variational approximation to CRPequivalent seqddCRP-mixture model, where it could be considered as alternative to one based on truncated stick-breaking representation. This allowed us to achieve significantly better variational lower bound than variational approximation based on truncated stick breaking for Dirichlet process.
Background: Gaussian processes (GP) provide an elegant and effective approach to learning in kernel machines. This approach leads to a highly interpretable model and allows using the Bayesian framework for model adaptation and incorporating the prior knowledge about the problem. The GP framework is successfully applied to regression, classification, and dimensionality reduction problems. Unfortunately, the standard methods for both GP-regression and GP-classification scale as O(n 3 ), where n is the size of the dataset, which makes them inapplicable to big data problems. A variety of methods have been proposed to overcome this limitation both for regression and classification problems. The most successful recent methods are based on the concept of inducing inputs. These methods reduce the computational complexity to O(nm2 ) where m is the number of inducing inputs with m typically much less than n. The present authors focus on classification. The current state-of-the-art method for this problem is based on stochastic optimization of an evidence lower bound (ELBO) that depends on O(m2 ) parameters. For complex problems, the required number of inducing points m is fairly big, making the optimization in this method challenging. Methods: The structure of variational lower bound that appears in inducing input GP classification has been analyzed. First, it has been noted that using quadratic approximation of several terms in this bound, it is possible to obtain analytical expressions for optimal values of most of the optimization parameters, thus sufficiently reducing the dimension of optimization space. Then, two methods have been provided for constructing necessary quadratic approximations: one is based on Jaakkola–Jordan bound for logistic function and the other is derived using Taylor expansion. Results: Two new variational lower bounds have been proposed for inducing input GP classification that depend on a number of parameters. Then, several methods have been suggested for optimization of these bounds and the resulting algorithms have been compared with the state-of-the-art approach based on stochastic optimization. Experiments on a bunch of classification datasets show that the new methods perform the same or better results than the existing one. However, new methods do not require any tunable parameters and can work in settings within a big range of n and m values, thus significantly simplifying training of GP classification models.
We propose a method (TT-GP) for approximate inference in Gaussian Process (GP) models. We build on previous scalable GP research including stochastic variational inference based on inducing inputs, kernel interpolation, and structure exploiting algebra. The key idea of our method is to use Tensor Train decomposition for variational parameters, which allows us to train GPs with billions of inducing inputs and achieve state-of-the-art results on several benchmarks. Further, our approach allows for training kernels based on deep neural networks without any modifications to the underlying GP model. A neural network learns a multidimensional embedding for the data, which is used by the GP to make the final prediction. We train GP and neural network parameters end-to-end without pretraining, through maximization of GP marginal likelihood. We show the efficiency of the proposed approach on several regression and classification benchmark datasets including MNIST, CIFAR-10, and Airline.
A model for organizing cargo transportation between two node stations connected by a railway line which contains a certain number of intermediate stations is considered. The movement of cargo is in one direction. Such a situation may occur, for example, if one of the node stations is located in a region which produce raw material for manufacturing industry located in another region, and there is another node station. The organization of freight traffic is performed by means of a number of technologies. These technologies determine the rules for taking on cargo at the initial node station, the rules of interaction between neighboring stations, as well as the rule of distribution of cargo to the final node stations. The process of cargo transportation is followed by the set rule of control. For such a model, one must determine possible modes of cargo transportation and describe their properties. This model is described by a finite-dimensional system of differential equations with nonlocal linear restrictions. The class of the solution satisfying nonlocal linear restrictions is extremely narrow. It results in the need for the “correct” extension of solutions of a system of differential equations to a class of quasi-solutions having the distinctive feature of gaps in a countable number of points. It was possible numerically using the Runge–Kutta method of the fourth order to build these quasi-solutions and determine their rate of growth. Let us note that in the technical plan the main complexity consisted in obtaining quasi-solutions satisfying the nonlocal linear restrictions. Furthermore, we investigated the dependence of quasi-solutions and, in particular, sizes of gaps (jumps) of solutions on a number of parameters of the model characterizing a rule of control, technologies for transportation of cargo and intensity of giving of cargo on a node station.
Event logs collected by modern information and technical systems usually contain enough data for automated process models discovery. A variety of algorithms was developed for process models discovery, conformance checking, log to model alignment, comparison of process models, etc., nevertheless a quick analysis of ad-hoc selected parts of a journal still have not get a full-fledged implementation. This paper describes an ROLAP-based method of multidimensional event logs storage for process mining. The result of the analysis of the journal is visualized as directed graph representing the union of all possible event sequences, ranked by their occurrence probability. Our implementation allows the analyst to discover process models for sublogs defined by ad-hoc selection of criteria and value of occurrence probability
The geographic information system (GIS) is based on the first and only Russian Imperial Census of 1897 and the First All-Union Census of the Soviet Union of 1926. The GIS features vector data (shapefiles) of allprovinces of the two states. For the 1897 census, there is information about linguistic, religious, and social estate groups. The part based on the 1926 census features nationality. Both shapefiles include information on gender, rural and urban population. The GIS allows for producing any necessary maps for individual studies of the period which require the administrative boundaries and demographic information.
It is well-known that the class of sets that can be computed by polynomial size circuits is equal to the class of sets that are polynomial time reducible to a sparse set. It is widely believed, but unfortunately up to now unproven, that there are sets in EXPNP, or even in EXP that are not computable by polynomial size circuits and hence are not reducible to a sparse set. In this paper we study this question in a more restricted setting: what is the computational complexity of sparse sets that are selfreducible? It follows from earlier work of Lozano and Torán (in: Mathematical systems theory, 1991) that EXPNP does not have sparse selfreducible hard sets. We define a natural version of selfreduction, tree-selfreducibility, and show that NEXP does not have sparse tree-selfreducible hard sets. We also construct an oracle relative to which all of EXP is reducible to a sparse tree-selfreducible set. These lower bounds are corollaries of more general results about the computational complexity of sparse sets that are selfreducible, and can be interpreted as super-polynomial circuit lower bounds for NEXP.