Nonparametric probabilistic models for learning spatiotemporal patterns of stream data

Abstract

The complexity in temporal domains requires introduction of the unobservable inherent dependencies. However, this task is very challenging due to difficulties such as hierarchical characteristics of the domain and abrupt changes in modalities. This thesis addresses the problem of extracting inherent dependencies in temporal domains. We develop algorithms for modeling and inference: inference with changing underlying distributions, temporal sequence learning with summarized problem space, and inference in a multichannel sequential environment. There exist domains where it is not possible to select a stable underlying distribution due to environmental causes. We develop an inference algorithm based on the feature relevance network in order to tackle the distribution change. The idea of feature relevance network is inspired by the existence of stable relations between features irrespective of environmental change. The proposed algorithm is verified on an indoor location estimation task. We show that the non-parametric approach making no assumption on feature distributions is capable of discovering hidden relations among features. The most obvious instances of temporal domain are dynamic streams such as TV drama. In order to estimate hidden dependencies in an episode of TV drama and represent a whole episode in a more succinct form, this thesis introduces a temporal learning scheme based on a set of particles. Instead of assuming a prior distribution, particles capture prominent characteristics of a given stream in a collaborative manner. The proposed method evolves particles through two-stage learning. At the first stage, a segment (scene) is estimated using evolutionary particle filtering (PF). At the second stage, a transitional probability matrix representing dependencies between estimated segments is computed and stored. We demonstrate performance by comparing to human-evaluated ground truth and regenerating images of expected sequences succeeding a given seed image. For inference in a multichannel sequential environment, the sequential hierarchical Dirichlet process (sHDP) is developed. For seamless processing of multichannel data, sHDP divides a multichannel stream into sub-channel of single modality and builds a latent model for a sub-channel. Changes in sub-channels are linked though a dynamic channel merging scheme. The proposed method is applied for semantic segmentation in real TV drama episodes and the performance is compared to human evaluated ground-truth. By incorporating inherent dependencies, we present successful algorithms to deal with applications in temporal domains. The three areas of this thesis are promising realizations of inherent dependencies learning, and the algorithms presented here inform the range of possibility of applicability in temporal domains.