[ Bristol CS | Index | ML group | Peter Flach | Papers | Presentations | Student projects ] A system developed at the ACRC Vision Group uses neural networks to identify images in an image database that bear a certain similarity to a user-selected region in an image. The system returns a number of thumbnails, which the user then classifies as positive and negative. The neural network is retrained on this new dataset, and the process iterates.
Symbolic learning systems learn definitions of the target concept consisting of classification rules. Classifying an unseen instance is done by reasoning with the learned rules. As a consequence, the system will be able to explain its classification. Also, the user can adjust feedback so as to correct erroneous rules.
This project concerns replacing the neural net learning engine of the image classification system with a symbolic learning engine (e.g. C4.5). A possible second goal of the project is to employ first-order rules (inductive logic programming) in order to describe multiple regions in a picture and their spatial relation (e.g. a white triangular region above a large blue region).