The emphasis of our research is on measuring and improving the performance of techniques for analysing image or spatial data. The techniques that we use include:
- Boltzmann Machines
- Generative Adversarial Training
- Reinforcement Learning
- Networks with pre-defined weights (see “Historical”)
In addition to our focus on improving techniques, we also work on applying these to a wide range of problems: our Publications page contains a mixture of applied and fundamental research.
Each year, we also run undergraduate or Masters’ student projects that are technology-oriented or application-oriented.
For almost two decades, we used Wavelet Representations as adaptive front-ends for the analysis of visual data. These are linear and overcomplete representations that are then subjected to non-linearities. The architecture and flow of processing bears strong resemblance to that of trained convolutional neural networks.
We have also developed custom descriptors for different types of problem in image analysis and computer vision, such as for colour processing.