| Network |
Inventor / Developer |
Years |
Applications |
Limitations |
Comments |
| Adaptive resonance theory |
Gail Carpenter, Stephen Grossberg |
1978-86 |
Pattern recognition, (radar / sonar, voiceprints) |
Sensitive to translation, distortion, scale |
Sophisticated; not applied much yet |
| Avalanche |
Stephen Grossberg |
1967 |
Continuous-speech recognition; commands to robot arms |
No easy way to alter speed or interpolate movement |
Class of networks; no single neural network can do all this |
| Back Propagation |
Paul Werbos, David Parker, David Rumelhart |
1974-85 |
Speech systhesis from text; robot arms; bank loans |
Supervised training only - needs lots of I/O examples |
Most popular neural network; works well, easy to learn, powerful |
| Bidirectional associative memory |
Bart Kosko |
1985 |
Content-addressable associative memory |
Low storage density; data must be coded |
Easiest to learn; associate fragment pairs with complete pairs |
| Boltzmann & Cauchy machines |
Jeffrey Hinton, Terry Sejnowsky, Harold Szu |
1985-86 |
Pattern recognition for images, sonar, radar |
Long training time, generates noise in statis distribution |
Simple neural network; noise function used to find global minima |
| Brain State in a Box |
James Anderson |
1977 |
Extraction of knowledge from data bases |
One-shot decisions - no iteration |
Similar to bidirection in completing fragmented inputs |
| Cerebellatron |
David Marr, James Albus, Andres Pellionez |
1969-82 |
Control motor action of robotic arms |
Requires complicated control input |
Like Avalanche; can blend commands with different weights for smoothness |
| Counter-propagation |
Robert Hecht-Nielsen |
1986 |
Image compress; stat analysis; loan application score |
Many PEs and connections for high accuracy |
Self-propagation look-up table; similar to back propagation |
| Hopfield |
John Hopfield |
1982 |
Retrieval of complete data from fragments |
Does not learn; weights must be set in advance |
Can implement on a large scale |
| MADALINE |
Bernard Widrow |
1960-62 |
Nulling of radar jammers; modems; phone equalisers |
Assume linear relationship between I & O |
In commercial use more than 20 years; powerful learning law |
| Neocognitron |
Kunihiko Fukushima |
1978-84 |
Handprinted character recognition |
Requires many PEs and connects |
Most complicated; insensitive to scale translation, rotation |
| Perceptron |
Frank Rosenblatt |
1957 |
Typed-character recognition |
Cannot recognise complex characters; sensitive to scale, distortion |
Oldest neural network; was built in hardware, rarely used today |
| Self-organising map |
Teuvo Kohonen |
1980 |
Maps 1 geometric region (grid) to another (aircraft) |
Require much training |
More effective than many algorithms for aerodynamic flow calculations |
