Data Neutronium. Phase 1
Abstract
This project is developing a system to compress data such as imagery and speech to near-maximum levels. We use Cottrell/Munro/Zipser neural networks to implement a vector quantization method (called data neutronium, in analogy with solid neutronium, the densest possible form of matter) that uses a mathematically defined codebook on the data manifold from which the data to be compressed is drawn. The C/M/Z neural network (which is trained off-line, once) is computationally simple and can carry out both data compression and decompression in real-time using low-cost hardware. In Phase I we demonstrated 64:1 compression of 8-bit per pixel imagery at an RMS pixel error of 20.7 grey scale levels (our goal was 50:1 with an RMS pixel error below 25). We also developed a mathematical proof that, for the case of large compression problems (e.g., large image tiles or long speech sample time windows), the mean squared error distortion of the data neutronium method will be no more than 43% greater than that of an optimal source coding system with the same number of code bits. Thus, data neutronium is the only potentially practical data compression method that is provably near-optimal.... Data Compression, Source Coding Theory, Neural Networks, Differential Topology, Image Compression, Information Theory.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 19, 1993
- Accession Number
- ADA263648
Entities
People
- Robert Hecht-neilson
- Shinmin Wang