Training and application of neural networks constructed from RAMs

Exhibitor: Robert Lee

Supervisor: Tom Downs

Research Group: Complex and Intelligent Systems

Industry Sector: Scientific / Research Services

In the existing technology, Artificial Neural Network is applied to many fields of application. Specifically enough, these types of networks are constructed in computing or mathematical basis. The use in relation to industries is generally noticeable. This includes medical, mining, gaming, finance, security, weather forecast, database, simple pattern recognition, and many more. Usually there is a sense of automation involved when neural network is applied; allowing computers to “learn” whatever information is supplied to their input. Then make use of the stored information by supplying the network with unseen input samples to generate an output.

In my proposal, the design of the network employs a combination of N-Tuple Neural Network (NTNN) and Multi-Layered Ram Net (MLRN) architecture. Minchinton Cells are employed for data encoding and compression. Lookup Tables (LUTs) are used to simulate hardware associative memories, the Correlation Matrix Memory.

The dataset used in this project is the Sonar Dataset created by Gorman and Sejnowski in 1988 to test hidden units in a layered network.

Specific to implementation, OpenGL engine is used to provide data analysis and network performance evaluation to create a sense of visualisation. The source code is written in ANSI-C style for possible future research and references.

For future improvements, more than one dataset should be used and a better file-reading algorithm should be applied for various conversions of the datasets.

-Robert Lee s363039@student.uq.edu.au

Thesis Document (PDF)

Other Related Files