This project considers the practical application of theoretical methods to access to dependability of neural networks. As a focus of our work a region classification problem has been adopted. To support this British Aerospace provided a database consisting of 119 colour images which have been segmented and the subsequent regions labelled. Using this database a set of 35 features was obtained and we trained a neural network to classify each region (object) as 1 of 11 classes (e.g. sky, building, vegetation, road, car).
The following figure shows what an image labelled by the network looks like. A different colour has been used to designate the identity of each region. The table on the right gives the colour/region type relationship.
Typical region labelling produced by a Neural Network. |
|
In addressing the dependability of this method, we have identified three main lines of research.
The approach has followed two main directions:
b) The Theoretical investigation of learning curves. In the limit of an infinite number of hidden nodes, neural networks can be interpreted as Gaussian Processes. we are using this approximation to study generalisation capabilities versus the amount of training data.
In any classification problem a number of features may be correlated or redundant. If the input vectors are assumed to lie on a low-dimensional manifold in a higher dimensional space then it may be possible to reduce the number of features, making the input coding to the network more efficient. There are a number of techniques for tackling this issue of features reduction. Some of them involve the step-wise deletion (or addition) of features in the input vector. However, there are more principled and sophisticated methods.
We have investigated this issue by following two approaches:
Project Home page | Introduction | Work plan | Publications |