On the Validity of Multi-Resolution Non-Supervised Training Algorithms

Speaker: Dr Dan Tamir

Time: 12:30pm-1:30pm, March 30th, 2007

Location: Nueces 201 Conference Room

Abstract:

A multi-resolution method for training the LBG vector quantizer and a Kohonen 
competitive neural network (KCNN) is presented. Starting with a low resolution sample 
of the input data, the training algorithm is applied to a sequence of 
monotonically increasing-resolution samples of the data. The final centers 
(weight matrix) obtained from a low resolution stage are used as the initial 
centers for the next stage which is a higher resolution stage. 
The multi-resolution algorithm is tested using synthetic data and several 
multi-spectral images and compared to the traditional LBG / KCNN. It is 
found that in the average case the multi-resolution reduces the computation 
time by a factor of more than four for LBG-VQ and two for KCNN-VQ with a 
slight improvement in the quality of quantization
The validity of the multi-resolution algorithm is assessed through evaluation 
of the entropy of the results, the entropy of the problem space, and the 
probability of achieving a global optimum. A discussion on alternative 
approaches for validation will conclude the presentation.