KEDRI - Centre for Adaptive Pattern Recognition
Centre for Adaptive Pattern Recognition
This page is a summary of what our centre does and who we are. It covers:
The centre develops novel methods for signal processing and modelling by using a combination of different modalities:
- Speech
- Image
- Video
- Motion
- Sensory information.
Current research projects
- Speech recognition
- Image and video processing
- Odour recognition (“artificial nose”)
- Adaptive modelling of human motion and application in sport couching
- Integrated multimodal systems (speech and image) (Fig. 1)
- Biometric applications
- Adaptive mobile robots (Fig.2)
Figure 1. An integrated environment for person identification system development, based on image and speech information, was patented as a PCT.
Figure 2. The robotic soccer system developed by L.Huang utilised the KEDRI on-line learning method ECF and won a gold medal at the World ROBOCUP 2004 in Seoul.
Hierarchy Minimum Enclosing Balls (HMEB) for Multi-label Classification
Project Team
- Prof. Nikola Kasabov
- Dr. Paul Pang
- Mr. Gary Chen
As network intrusions are usually associated with multiple labels in a hierarchy structure, the classification of network intrusions naturally possess a hierarchical multi-label classification (HMC) problem, in which every instance may belong to more than one class. Addressing HMC problem, we propose a novel approach of hierarchy minimum enclosing balls (HMEB), where a label hierarchy is modeled as data HyperSphere hierarchies, called minimum enclosing ball (MEB). The MEBs are separating, encompassing and overlapping with each other formed as tree-like structure.
Given an unlabeled sample, HMEB seeks a MEB enclosing the sample, and label the sample according to the MEB's position in the MEB hierarchy. The proposed method has been validated on the hierarchy multi-label problem from KDD'99 and Yeast dataset, respectively. The experimental results show that the proposed HMEB improves the classification accuracy of U2R from 13.2% to 82.7% and R2L from 8.4\% to 45.9%, as compared to the winner of KDD'99. Also, the efficiency of HMEB is highlighted, as the computational time stays steady while the size of training data exponentially manifolds in another experiment of Yeast data.
Adaptive modelling of human motion and a methodology of the discovery of coaching rules
Project Team
- Mr. Boris Bacic
- Mr. David H. Zhang
- Dr Zeke Chan
- Prof. N.Kasabov
- Prof. S.MacDonell
The aim of this research is to assess the suitability of evolving connectionist systems (ECOS) to extract tennis coaching rules from biomechanical time-series data [1]. We show that a prototype system is able to provide feedback obtained from extracted and transformed fuzzy rules, and is adaptable to previously unseen data.
A sport coaching is essentially the provision of personalised feedback that progresses from resolving fundamental problems (for novice players) through to more detailed and specific advice (for expert players). In producing an automated tennis coach these principles need to be retained. The prototype solution described here observes some of these key factors (i.e. attributes derived from hand motion trajectory relative to the body motion), considering first the elementary errors made by novice learners that in turn cause subsequent errors in tennis technique. After classification of each learner’s tennis swing personalised feedback can be produced from extracted coaching rules using ECOS tools. As the feedback from different tennis coaches may vary and come in fuzzy form the prototype would also need to provide fuzzy feedback. To produce even more human-intuitive feedback obtained from "if-then" extracted fuzzy rules, they need to be further translated and the hierarchy of fuzzy rules (from elementary to sophisticated) explored. Two system architectures (Fig. 1) are candidates for such a system. Both can accept multiple sets of features as input and provide structured rules based on intermediate information from hierarchical subsystem output or from fuzzy layer rules within a single evolving neuro-fuzzy system.
Fig 1. Proposed structures of subsystems as single neuro-fuzzy architecture.
As an input to the proposed architectures different feature subsets (i.e. key factors contributing to a good or bad swing) have been computed from biomechanical data. Processing these feature sets into common output classes (i.e. good or bad swing) using EFuNN and ECF, it is possible to extract knowledge as fuzzy coaching rules and transform this knowledge further with the goal of making it more accessible to tennis players.
Our researchers
- Professor Nikola Kasabov, Center Director
- Dr Stephan Schliebs
- Dr Ammar Mohemmed
- Gary Chen
- Kshitij Dhoble
- Marin Karaivanov
- Boris Bacic
- Haza Nuzly
- Professor Alvis Fong (Associate Researcher)
