Using Recurrent Networks for Dimensionality Reduction
Abstract
This thesis explores how recurrent neural networks can be exploited for learning certain high-dimensional mappings. Recurrent networks are shown to be as powerful as Turing machines in terms of the class of functions they can compute. Given this computational power, a natural question to ask is how recurrent networks can be used to simplify the problem of learning from examples. Some researchers have proposed using recurrent networks for learning fixed point mappings that can also be learned on a feedforward network even though learning algorithms for recurrent networks are more complex. An important question is whether recurrent networks provide an advantage over feedforward networks for such learning tasks. The main problem with learning high- dimensional functions is the curse of dimensionality which roughly states that the number of examples needed to learn a function increases exponentially with input dimension. Reducing the dimensionality of the function being learned is therefore extremely advantageous. This thesis proposes a way of avoiding the curse of dimensionality for some problems by using a recurrent network to decompose a high-dimensional function into many lower dimensional functions connected in a feedback loop and then iterating to approximate the high- dimensional function. This idea is then tested on learning a simple image segmentation algorithm given examples of segmented and unsegmented images.... Neural networks, Recurrent networks, Image segmentation, Dimensionality reduction.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 01, 1992
- Accession Number
- ADA259497
Entities
People
- Michael J. Jones
Organizations
- Massachusetts Institute of Technology