Nearly a Neural Net

Kevin Blithe, Anita Bowles, Noelle LaVoie, Bryan Loughry

The "Nearly a Neural Net" project is composed of: a set of nine input switches, nine Crickets serving as nodes in the network, and nine output lights. The network is mounted on a black wooden frame with the switch and output light wires hidden on the underside. The input switches are black rocker switches painted white on the underside so that the user can easily see when the switches are on and when they are off. The nodes are red dot Crickets programmed to communicate their activation values to one another via infrared. The output lights are nin volt miniature incandescent bulbs purchased from a local electrical supply company.

The user interacts with the project by inputting one of two partially orthogonal (only one input is shared by both) patterns on which the network has been pretrained. The patterns are input by flipping the rocker switches. After the pattern is completed, the user presses the "Think" button which initiates network activity. The pattern computed by the network is then displayed on the nine output lights. In addition, after testing the learned patterns, users are encouraged to input a variety of partial patterns in order to test the robustness of the network representation. Accompanying text and graphics explain the basics of neural networks and the use of top-down processing in visual perception.

preliminary project sketch