Validation and Verification of Neural Network Systems
Work plan
The time scale of the project is as follows:
Year 1
During the first year we have implemented a
Bayesian training of neural networks
to classify outdoors images.
Within this framework we have also addressed the issue of error bars
on prediction and feature selection.
Year 2
A further investigation on the issue of
feature selection has been achieved during the second year
through the implementation
of
Automatic Relevance Determination.
We have also started the empirical investigation of the learning
curves.
Year 3
The final part of this research concerns more theoretical issues.
In particular we investigate the issue of generalisation
from a theoretical point of view, trying to understand how
generalisation is affected by the amount of training data.
The study of the issue of dimensionality reduction
using latent variables has already started.