Related Experiment Videos
Randomly connected sigma-pi neurons can form associator networks
1Bios Group Inc., Santa Fe, NM 87501, USA. tony.plate@biosgroup.com
Network (Bristol, England)
|December 29, 2000
Summary
This study introduces a novel neural network model where information is stored in activity patterns, not connection strengths. This approach offers a plausible mechanism for cognitive functions like analogy processing and evolutionary origins of neural systems.
Area of Science:
- Computational Neuroscience
- Artificial Intelligence
- Cognitive Science
Background:
- Traditional associative memories rely on connection strengths to store information.
- Higher-level cognitive tasks require information to be accessible for further processing.
Purpose of the Study:
- To propose a new type of hetero-associative memory network using sigma-pi units.
- To demonstrate how information can be encoded in activity patterns rather than connection weights.
- To explore the implications for cognitive functions and evolutionary plausibility.
Main Methods:
- Utilized sigma-pi units randomly connected to two input vectors to form an encoding network.
- Developed a decoding network whose connectivity depends on the encoding network.
- Investigated the learning process for the decoding network's connectivity.
Main Results:
- Associations are represented as patterns of activity, not connection strengths.
- Information encoded in activation values is accessible for further processing.
- The proposed network demonstrates potential for tasks like analogy processing.
Conclusions:
- This model offers a biologically plausible mechanism for associative memory.
- Encoding information in activity patterns facilitates higher-level cognitive functions.
- Randomly connected networks can perform complex operations, supporting evolutionary theories of neural systems.