| Summary: | Recursive Bayesian classification (RBC) requires optimal latent variable estimation in the presence of noisy observation to achieve real-time sequential decision making. Active RBC introduced in this dissertation attempts to effectively select queries that lead to more informative observations to rapidly reduce uncertainty until a confident decision is made. Accordingly, active RBC includes the following fundamental components:(S)A stopping criterion based on the posterior
probability to stop evidence collection;(Q)a querying step to decide how to collect further evidence from relevant sources to benefit speed and accuracy objectives of RBC;(C)a classification objective based on the posterior distribution and loss values attributed to each true label and decision option pair to determine the optimal decision once the stopping criterion has been satisfied. This dissertation specifically focuses on optimizing querying (Q) and stopping (S) for RBC.
Conventional stopping criterion design methodologies lack insight of the RBC geometry and evolution of the posterior probability vector. Additionally, conventional active querying methods stagger due to misleading prior information. In this case, the system uses time inefficiently to overcome the provided belief by querying most likely candidates a number of iterations. Furthermore, in contrast to inference and querying being coactive, typically the optimality objectives are designed
separately. An electroencephalography (EEG)-based brain computer interface (BCI) system specifically de-signed for typing is used as a testbed for active RBC. BCI systems provide a communication pathway between the user and the environment both in medical and non-medical domains. EEG signals are widely used with promising performance to estimate user intent in BCI systems. BCI typing systems are epitomes of RBC driven systems as repeated evidence collection is mandated due to highly
variable EEG signals given a particular user intent (latent variable hidden in the brain). However, in many cases, EEG-based communication staggers and lacks accuracy and speed due to inefficient RBC. To increase the performance of RBC, motivated by information theoretic approaches to coding and active learning this dissertation contributes to the literature in three folds: (i) A complete analysis of stopping criterion and geometrical description of the RBC problem is provided.
Motivated by the posterior motion a stopping criterion design is proposed. Moreover, an early stopping scheme with one step ahead prediction is shown to make a decision with marginal accuracy deficit. (ii)Influenced by the posterior motion, a new query selection objective is proposed. This querying mechanism is shown to result in rapid and accurate inference in scenarios in which the recursive inference starts with a misleading (or adversarial) prior probability distribution for the
latent variable of interest (e.g. user attempting to type a letter/word that is unlikely according to the language model). (iii) Querying and stopping approaches are taken together into consideration and an experimental study specifically on BCI typing is presented. Additionally, the dissertation shows it is possible to reformulate RBC with Rényi entropy measures solidifying the connection between stopping and querying objective design. All contributions are verified using a BCI typing
system "BCIPy" with simulations and human-in-the-loop experiments.
|
|---|
