Recent developments in computational intelligence have resulted in a basic understanding of information processing and computing paradigms in biological systems and the nervous system in particular. A fundamental characteristic of these systems is the ability to represent and process a large amount of knowledge efficiently and with apparent ease. Research into brain-like systems is motivated by the desire to gain a further understanding of their information encoding and computing principles, and to reveal the mechanisms that underlie their adaptive and learning capabilities. The ultimate goal is to exploit these findings in building new bio-inspired systems that outperform traditional algorithmic and machine learning-based systems. Studies of the computational principles at the neuron level have resulted in a variety of neural models which capture the essential dynamics of biological neurons such as the use of spikes or action potential as a means of information representation and processing. These models have been used as a building block for different neural architectures with the aim of understanding the underlying mechanisms of neural signalling, and computing and learning in neural circuitries. So far, the lack of efficient learning algorithms has, however, hampered the applicability of these models and has restricted their use in developing efficient and scalable computational systems that can outperform classical approaches in solving real world problems such as those encountered in computer vision, robotics, speech recognition and prediction.
This project aims to develop new supervised and unsupervised learning approaches, using different neural models and architectures, and create new adaptive autonomous systems. The devised approaches will be evaluated on a range of benchmark datasets and real world applications such as image recognition used in content based image retrieval (CBIR), automatic face and speech recognition, and prediction.
Rationale / Context for the project
The project will build on previous and related PhD projects ongoing in the research centre.
The project will study the existing biologically plausible neural models, compare and evaluate the variety of existing learning algorithms and their applicability. The challenges of developing efficient learning algorithms for different neuron models and topologies will then be identified and addressed. The project will design and implement new learning algorithms and seek novel applications in image recognition and CBIR, face and speech recognition, financial time-series prediction and robotics. The performance of the devised approaches will also be evaluated and compared against existing approaches.
It is anticipated that this work will develop new neural learning paradigms and adaptive brain-like systems which will be demonstrated on real world applications. Results of this research will be disseminated in international conferences and journals.
- The researcher will require access to MATLAB suite for rapid prototyping and validations of developed algorithms. The following toolboxes are required: Neural networks, statistics, image processing and signal processing, the freely available such as Liquid States Machine (LSM) and Support Vector Machines (SVM) toolboxes.
- Benchmark datasets and a camera/acquisition card are required for computer vision applications.
- Benchmark speech corpora for speech application.
First Supervisor: Belatreche, A Dr
Second Supervisor: Li, Y Dr
Third Supervisor: Maguire, L Prof
Collaboration: This project does not involve collaboration with another establishment