Background
Real-world environments are often non-stationary. Learning systems such as classifiers used for pattern recognition, designed using even advanced learning methods (e.g. SVMs) make the stationarity assumption in input data distribution density during training and test or operation phases. As a result, their field/generalisation performance, at times, becomes unsatisfactory and therefore such learning systems may have only limited practical use. A very good example of a non-stationary environment is an EEG-based brain-computer interface (BCI) in which a communication link is established between a computer and the user through a process of simply thinking or imagining a set of cognitive tasks [1, 2]. In this, a classifier is used to detect the electrophysiological correlates appearing in the non-stationary EEG, resulting from the user’s cognitive tasks. The non-stationary characteristics of EEG has been one of the primary impediments in developing a practically useful BCI, despite tremendous R&D over the last two decades or so. In fact prevalence of such non-stationary environments is wide-spread starting from the autonomous robots operation in unstructured environment to share price trend predictions in share markets. Among other specific example applications, surveillance [5, 6], and spam filtering [7] show non-stationary characteristics, which is also called the covariate shift phenomenon [2, 3, 4, 5, 6]. In all these areas, a learning system devised under the assumption of input data distribution invariance results in suboptimal performance while the system is being used either as a classifier or a predictive model. At ISRC, we have developed promising adaptive learning algorithms and applied for accounting non-linearity and variability in EEG-based BCI. A self-organising fuzzy neural network (SOFFN) based learning technique [1] has been applied for pre-processing of EEG signal in a supervised mode. A linear discriminant analysis (LDA) based co-variate shift minimisation (CSM) technique [2] has been applied as a classifier to account for feature non-stationarity in a two-class BCI with promising results. However, the CSM technique is devised for a two-class system under certain limiting assumptions. Yet another related work is Lattice Machine [8] that uses hyperrelations to represent training data, and uses contextual probability [9] for classification. Non-stationarity can be accounted for through the variability in hyperrelations. Building on these promising works, this project aims to investigate how non-stationarities can be effectively accounted for in an automated way, in a wide-range of evolving systems.
Research Program
To account for the covariate shift, a learning system needs to take into consideration certain critical aspects: (1) It should learn from the changes in inputs alone; (2) There is a need to devise adaptation strategies that can make effective use of the available information to extract knowledge about the system variability and also make continuous update to account for the new information; and (3) The on-line update should be done only when the system variability has been detected but only when the variability is not because of noise. In the light of these, the project will involve undertaking a comprehensive review of important existing approaches and then investigate whether a superior hybrid method can be devised by combining novel features taken from multiple methods. A highly probable research direction may involve estimating system non-stationarity through the co-variate shift estimation. Based on the co-variate shift estimation, one of the following two tasks can be used for the classifier adaptation: Update or shift the class separation plane/surface based on the covariate shift; and update the classifier parameters based on the covariate shift.
A major focus will be on thorough evaluation of the method in multiple application areas. This will help assess whether application specific novelties can be exploited in the algorithm design and also whether a universally applicable algorithm can be designed to achieve the best performance in any system irrespective of its specific characteristics.
Anticipated outcomes
It is anticipated that a successful completion of the project will result into a practical adaptive learning algorithm with wide-ranging applicability in non-stationary systems.
References
Resources Needed: All resources are available.
First Supervisor: Wang, H Dr
Second Supervisor: Prasad, G Dr
Collaboration: This project does not involve collaboration with another establishment
Real-world environments are often non-stationary. Learning systems such as classifiers used for pattern recognition make the stationarity assumption in input data distribution density during training and test or operation phases. As a result, their field/generalisation performance, at times, becomes unsatisfactory and therefore such learning systems may have only limited practical use. This project aims to investigate how non-stationarities can be effectively accounted for in an automated way, in a wide-range of evolving systems.
A major focus will be on thorough evaluation of the method in multiple application areas. This will help assess whether application specific novelties can be exploited in the algorithm design and also whether a universally applicable algorithm can be designed to achieve the best performance in any system irrespective of its specific characteristics.